We introduce LM-Lexicon, an innovative definition modeling approach that incorporates data clustering, semantic expert learning, and model merging using a sparse mixture-of-experts architecture. By decomposing the definition modeling task into specialized semantic domains, where small language models are trained as domain experts, LM-Lexicon achieves substantial improvements (+7% BLEU score compared with the prior state-of-the-art model) over existing methods on five widely used benchmarks. Empirically, we demonstrate that 1) the clustering strategy enables fine-grained expert specialization with nearly 10% improvement in definition quality; 2) the semantic-aware domain-level routing mechanism achieves higher expert efficacy (+1%) than conventional token-level routing; and 3) further performance gains can be obtained through test-time compute and semantic expert scaling. Our work advances definition modeling while providing insights into the development of efficient language models for semantic-intensive applications.

LLM agents have become increasingly sophisticated, especially in the realm of cybersecurity. Researchers have shown that LLM agents can exploit real-world vulnerabilities when given a description of the vulnerability and toy capture-the-flag problems. However, these agents still perform poorly on real-world vulnerabilities that are unknown to the agent ahead of time (zero-day vulnerabilities).In this work, we show that teams of LLM agents can exploit real-world, zero-day vulnerabilities. Prior agents struggle with exploring many different vulnerabilities and long-range planning when used alone. To resolve this, we introduce HPTSA, a system of agents with a planning agent that can launch subagents. The planning agent explores the system and determines which subagents to call, resolving long-term planning issues when trying different vulnerabilities. We construct a benchmark of 14 real-world vulnerabilities and show that our team of agents improve over prior agent frameworks by up to 4.3×.

Variation in human annotation (i.e., disagreements) is common in NLP, often reflecting important information like task subjectivity and sample ambiguity. Modeling this variation is important for applications that are sensitive to such information. Although RLVR-style reasoning (Reinforcement Learning with Verifiable Rewards) has improved Large Language Model (LLM) performance on many tasks, it remains unclear whether such reasoning enables LLMs to capture informative variation in human annotation. In this work, we evaluate the influence of different reasoning settings on LLM disagreement modeling. We systematically evaluate each reasoning setting across model sizes, distribution expression methods, and steering methods, resulting in 60 experimental setups across 3 tasks. Surprisingly, our results show that RLVR-style reasoning degrades performance in disagreement modeling, while naive Chain-of-Thought (CoT) reasoning improves the performance of RLHF LLMs (RL from human feedback). These findings underscore the potential risk of replacing human annotators with reasoning LLMs, especially when disagreements are important.

Quality estimation is omnipresent in machine translation, for both evaluation and generation. Unfortunately, quality estimation models are often opaque and computationally expensive, making them impractical to be part of large-scale pipelines. In this work, we tackle two connected challenges: (1) reducing the cost of quality estimation at scale, and (2) developing an inexpensive uncertainty estimation method for quality estimation. To address the latter, we introduce Instant Confidence COMET, an uncertainty-aware quality estimation model that matches the performance of previous approaches at a fraction of their costs. We extend this to Early-Exit COMET, a quality estimation model that can compute quality scores and associated confidences already at early model layers, allowing us to early-exit computations and reduce evaluation costs. We also apply our model to machine translation reranking. We combine Early-Exit COMET with an upper confidence bound bandit algorithm to find the best candidate from a large pool without having to run the full evaluation model on all candidates. In both cases (evaluation and reranking) our methods reduce the required compute by 50% with very little degradation in performance. Finally, we show how Instant Confidence COMET can be used to decide which translations a human evaluator should score rather than relying on the COMET score.

Decomposition-based multi-hop retrieval methods rely on many autoregressive steps to break down complex queries, which breaks end-to-end differentiability and is computationally expensive. Decomposition-free methods tackle this, but current approaches struggle with longer multi-hop problems and generalization to out-of-distribution data. To address these challenges, we introduce GRITHopper-7B, a novel multi-hop dense retrieval model that achieves state-of-the-art performance on both in-distribution and out-of-distribution benchmarks. GRITHopper-7B combines generative and representational instruction tuning by integrating causal language modeling with dense retrieval training. Through controlled studies, we find that incorporating additional context after the retrieval process, referred to as post-retrieval language modeling, enhances dense retrieval performance. By including elements such as final answers during training, the model learns to better contextualize and retrieve relevant information. GRITHopper-7B offers a robust, scalable, and generalizable solution for multi-hop dense retrieval, and we release it to the community for future research and applications requiring complex reasoning and retrieval capabilities.

Understanding long-context visual information remains a fundamental challenge for vision-language models, particularly in agentic tasks such as GUI control and web navigation. While web pages and GUI environments are inherently structured documents, current VLMs typically neglect decision-oriented document understanding in their training objectives. Existing approaches primarily extend visual embeddings to process long, high-resolution inputs, but these methods are memory-intensive and impractical for locally deployable solutions. To address these issues, we propose SCoPE VLM, a document navigation expert that leverages a novel Chain of Scroll mechanism to selectively and recursively navigate documents, focusing exclusively on relevant segments. We introduce a dedicated data generation pipeline to construct informative Chain of Scroll trajectories and Episodic Group Relative Policy Optimization, a tailored reinforcement learning method to bridge the gap between training and inference. Our method substantially reduces memory usage and effectively models human-like reading behaviors. To the best of our knowledge, SCoPE VLM is the first framework to explicitly model agentic reading patterns in multi-page document question answering, advancing the capabilities of multimodal agents.

High-quality instruction-tuning data is crucial for developing Large Language Models (LLMs) that can effectively navigate real-world tasks and follow human instructions. While synthetic data generation offers a scalable approach for creating such datasets, it imposes a quality ceiling where models trained on the data cannot outperform the LLM generating it. To overcome this limitation, we introduce Reference-Level Feedback, a paradigm that extracts desirable characteristics from carefully curated reference samples to guide the synthesis of higher-quality instruction-response pairs. Using this approach, we synthesize REFED, a dataset of 10K instruction-response pairs. Fine-tuning Llama-3.1-8B-Instruct and Mistral-7B-Instruct on REFED demonstrate state-of-the-art performance among similarly sized models, notably reaching a 43.96% length-controlled win-rate on AlpacaEval 2.0. Extensive experiments demonstrate that Reference-Level Feedback consistently outperforms traditional sample-level feedback methods, generalizes across model architectures, and produces high-quality and diverse data at low cost.

Since many real-world documents combine textual and tabular data, robust Retrieval Augmented Generation (RAG) systems are essential for effectively accessing and analyzing such content to support complex reasoning tasks. Therefore, this paper introduces T²-RAGBench, a benchmark comprising 23,088 question-context-answer triples, designed to evaluate RAG methods on real-world text-and-table data. Unlike typical QA datasets that operate under Oracle Context settings, T²-RAGBench challenges models to first retrieve the correct context before conducting numerical reasoning. Existing QA datasets containing text-and-table data typically contain context-dependent questions, which may yield multiple correct answers depending on the provided context. To address this, we transform SOTA datasets into a context-independent format, validated by experts as 91.3% context-independent questions, enabling reliable RAG evaluation. Our comprehensive evaluation identifies Hybrid BM25 , a technique that combines dense and sparse vectors, as the most effective approach for text-and-table data. However, results demonstrate that T²-RAGBench remains challenging even for SOTA LLMs and RAG methods. Further ablation studies examine the impact of embedding models and corpus size on retrieval performance. T²-RAGBench provides a realistic and rigorous benchmark for existing RAG methods on text-and-table data. Code and dataset are available online: https://github.com/uhh-hcds/g4kmu-paper

Current large language models (LLMs) have demonstrated emerging capabilities in social intelligence tasks, including implicature resolution and theory-of-mind reasoning, both of which require substantial pragmatic understanding. However, how LLMs acquire this pragmatic competence throughout the training process remains poorly understood. In this work, we introduce ALTPRAG, a dataset grounded in the pragmatic concept of alternatives, to evaluate whether LLMs at different training stages can accurately infer nuanced speaker intentions. Each instance pairs two equally plausible yet pragmatically divergent continuations and requires the model to (i) infer the speaker’s intended meaning and (ii) explain when and why a speaker would choose one utterance over its alternative, thus directly probing pragmatic competence through contrastive reasoning. We systematically evaluate 22 LLMs across three key training stages: after pre-training, supervised fine-tuning (SFT), and preference optimization, to examine the development of pragmatic competence. Our results show that even base models exhibit notable sensitivity to pragmatic cues, which improves consistently with increases in model and data scale. Additionally, SFT and RLHF contribute further gains, particularly in cognitive-pragmatic scenarios. These findings highlight pragmatic competence as an emergent and compositional property of LLM training and offer new insights for aligning models with human communicative norms.

Hierarchical text classification (HTC) depends on taxonomies that organize labels into structured hierarchies. However, many real-world taxonomies introduce ambiguities, such as identical leaf names under similar parent nodes, which prevent language models (LMs) from learning clear decision boundaries. In this paper, we present TaxMorph, a framework that uses large language models (LLMs) to transform entire taxonomies through operations such as renaming, merging, splitting, and reordering. Unlike prior work, our method revises the full hierarchy to better match the semantics encoded by LMs. Experiments across three HTC benchmarks show that LLM-refined taxonomies consistently outperform human-curated ones in various settings up to +2.9pp. in F1. To better understand these improvements, we compare how well LMs can assign leaf nodes to parent nodes and vice versa across human-curated and LLM-refined taxonomies. We find that human-curated taxonomies lead to more easily separable clusters in embedding space. However, the LLM-refined taxonomies align more closely with the model’s actual confusion patterns during classification. In other words, even though they are harder to separate, they better reflect the model’s inductive biases. These findings suggest that LLM-guided refinement creates taxonomies that are more compatible with how models learn, improving HTC performance.

In-Context Learning (ICL) emerges as a key feature for Large Language Models (LLMs), allowing them to adapt to new tasks by leverageing task-specific examples without updating model parameters. However, ICL faces challenges with increasing numbers of examples due to performance degradation and quadratic computational costs. In this paper, we propose Logit Arithmetic Reweighting Approach (LARA), a novel framework that enhances ICL by using logit-based ensembling of multiple demonstrations. Our approach divides long input demonstrations into parallelizable shorter inputs to significantly reduce memory requirements, and then effectively aggregate the information by reweighting logits of each group via a non-gradient optimization approach. We further introduce Binary LARA (B-LARA), a variant that constrains weights to binary values to simplify the search space and reduces memory usage by filtering out less informative demonstration groups. Experiments on BBH and MMLU demonstrate that LARA and B-LARA outperform all baseline methods in both accuracy and memory efficiency. We also conduct extensive analysis to show that LARA generalizes well to scenarios of varying numbers of examples from limited to many-shot demonstrations.

Conversational large language models are trained to refuse to answer harmful questions. However, emergent jailbreaking techniques can still elicit unsafe outputs, presenting an ongoing challenge for model alignment. This paper aims to deepen our understanding of how different jailbreak types circumvent safeguards by analyzing model activations on different jailbreak inputs. We find that it is possible to extract a jailbreak vector from a single class of jailbreaks that works to mitigate jailbreak effectiveness from other, semantically-dissimilar classes. This suggests that diverse jailbreaks may exploit a common internal mechanism. We investigate a potential common mechanism of harmfulness feature suppression, and find evidence that effective jailbreaks noticeably reduce a model’s perception of prompt harmfulness. These insights pave the way for developing more robust jailbreak countermeasures and lay the groundwork for a deeper, mechanistic understanding of jailbreak dynamics in language models.

Despite the impressive performance of Retrieval-augmented Generation (RAG) systems across various NLP benchmarks, their robustness in handling real-world user-LLM interaction queries remains largely underexplored. This presents a critical gap for practical deployment, where user queries exhibit greater linguistic variations and can trigger cascading errors across interdependent RAG components. In this work, we systematically analyze how varying four linguistic dimensions (formality, readability, politeness, and grammatical correctness) impact RAG performance. We evaluate two retrieval models and nine LLMs, ranging from 3 to 72 billion parameters, across four information-seeking Question Answering (QA) datasets. Our results reveal that linguistic reformulations significantly impact both retrieval and generation stages, leading to a relative performance drop of up to 40.41% in Recall@5 scores for less formal queries and 38.86% in answer match scores for queries containing grammatical errors. Notably, RAG systems exhibit greater sensitivity to such variations compared to LLM-only generations, highlighting their vulnerability to error propagation due to linguistic shifts. These findings highlight the need for improved robustness techniques to enhance reliability in diverse user interactions.

Public opinion surveys show cross-cultural differences in political opinions between socio-cultural contexts. However, there is no clear evidence whether these differences translate to cross-lingual differences in multilingual large language models (MLLMs). We analyze whether opinions transfer between languages or whether there are separate opinions for each language in MLLMs of various sizes across five Western languages. We evaluate MLLMs’ opinions by prompting them to report their (dis)agreement with political statements from voting advice applications. To better understand the interaction between languages in the models, we evaluate them both before and after aligning them with more left or right views using direct preference optimization and English alignment data only. Our findings reveal that unaligned models show only very few significant cross-lingual differences in the political opinions they reflect. The political alignment shifts opinions almost uniformly across all five languages. We conclude that in Western language contexts, political opinions transfer between languages, demonstrating the challenges in achieving explicit socio-linguistic, cultural, and political alignment of MLLMs.

Long-term memory is one of the key factors influencing the reasoning capabilities of Large Language Model Agents (LLM Agents). Incorporating a memory mechanism that effectively integrates past interactions can significantly enhance decision-making and contextual coherence of LLM Agents. While recent works have made progress in memory storage and retrieval, such as encoding memory into dense vectors for similarity-based search or organizing knowledge in the form of graph, these approaches often fall short in structured memory organization and efficient retrieval. To address these limitations, we propose a Hierarchical Memory Architecture that organizes and updates memory in a multi-level fashion based on the degree of semantic abstraction. Each memory vector at a higher level is embedded with a positional index encoding pointing to its semantically related sub-memories in the next layer. During the reasoning phase, an index-based routing mechanism enables efficient, layer-by-layer retrieval without performing exhaustive similarity computations. We evaluate our method on five task settings from the LoCoMo dataset. Experimental results show that our approach consistently outperforms five baseline methods, demonstrating its effectiveness in long-term dialogue scenarios.

The abundance of multimodal news in digital form has intensified demand for systems that condense articles and images into concise, faithful digests. Yet most approaches simply conduct unimodal text summarization and attach the most-similar images with the text summary, which leads to redundancy both in processing visual content as well as in selection of images to complement the summary. We propose MULSUM, a two-step framework: (i) a Cross-Vis Aligner that projects image-level embeddings into a shared space and conditions a pre-trained LLM decoder to generate a visually informed text summary, and (ii) a Diversity-Aware Image Selector that, after the summary is produced, maximizes images-relevance to the summary while enforcing pairwise image diversity, yielding a compact, complementary image set. Experimental results on the benchmark MSMO (Multimodal Summarization with Multimodal Output) corpus show that MULSUM consistently outperforms strong baselines on automatic metrics such as ROUGE, while qualitative inspection shows that selected images act as explanatory evidence rather than ornamental add-ons. Human evaluation results shows that our diverse set of selected images was 13% more helpful than mere similarity-based image selection.

This work presents a comprehensive evaluation of how quantization affects model bias, with particular attention to its impact on individual demographic subgroups.We focus on weight and activation quantization strategies and examine their effects across a broad range of bias types, including stereotypes, fairness, toxicity, and sentiment.We employ both probability- and generated text-based metrics across 13 benchmarks and evaluate models that differ in architecture family and reasoning ability.Our findings show that quantization has a nuanced impact on bias: while it can reduce model toxicity and does not significantly impact sentiment, it tends to slightly increase stereotypes and unfairness in generative tasks, especially under aggressive compression.These trends are generally consistent across demographic categories and subgroups, and model types, although their magnitude depends on the specific setting.Overall, our results highlight the importance of carefully balancing efficiency and ethical considerations when applying quantization in practice.

Conditional acceptability refers to how plausible a conditional statement is perceived to be. It plays an important role in communication and reasoning, as it influences how individuals interpret implications, assess arguments, and make decisions based on hypothetical scenarios. When humans evaluate how acceptable a conditional "If A, then B" is, their judgments are influenced by two main factors: the conditional probability of B given A, and the semantic relevance of the antecedent A given the consequent B (i.e., whether A meaningfully supports B). While prior work has examined how large language models (LLMs) draw inferences about conditional statements, it remains unclear how these models judge the acceptability of such statements. To address this gap, we present a comprehensive study of LLMs’ conditional acceptability judgments across different model families, sizes, and prompting strategies. Using linear mixed-effects models and ANOVA tests, we find that models are sensitive to both conditional probability and semantic relevance—though to varying degrees depending on architecture and prompting style. A comparison with human data reveals that while LLMs incorporate probabilistic and semantic cues, they do so less consistently than humans. Notably, larger models do not necessarily align more closely with human judgments.

When language models correctly parse "The cat that the dog chased meowed,” are they analyzing syntax or simply familiar with dogs chasing cats? Despite extensive benchmarking, we lack methods to distinguish structural understanding from semantic pattern matching. We introduce **CenterBench**, a dataset of 9,720 comprehension questions on center-embedded sentences (like "The cat [that the dog chased] meowed”) where relative clauses nest recursively, creating processing demands from simple to deeply nested structures. Each sentence has a syntactically identical but semantically implausible counterpart (e.g., mailmen prescribe medicine, doctors deliver mail) and six comprehension questions testing surface understanding, syntactic dependencies, and causal reasoning. Testing six models reveals that performance gaps between plausible and implausible sentences widen systematically with complexity, with models showing median gaps up to 26.8 percentage points, quantifying when they abandon structural analysis for semantic associations. Notably, semantic plausibility harms performance on questions about resulting actions, where following causal relationships matters more than semantic coherence. Reasoning models improve accuracy but their traces show semantic shortcuts, overthinking, and answer refusal. Unlike models whose plausibility advantage systematically widens with complexity, humans shows variable semantic effects. CenterBench provides the first framework to identify when models shift from structural analysis to pattern matching.

Antimicrobial resistance is a growing global health threat, driving interest in nanoparticle-based alternatives to conventional antibiotics. Inorganic nanoparticles (NPs) with intrinsic antibacterial properties show significant promise; however, efficiently identifying relevant studies from the rapidly expanding literature remains a major challenge. This step is crucial for enabling computational approaches that aim to model and predict NP efficacy based on physicochemical and structural features. In this study, we explore the effectiveness of traditional machine learning and deep learning methods in classifying scientific abstracts in the domain of NP-based antimicrobial research. We introduce the “Antibacterial Inorganic NAnoparticles Dataset” AINA of 7,910 articles, curated to distinguish intrinsic antibacterial NPs from studies focusing on drug carriers or surface-bound applications. Our comparative evaluation shows that a fine-tuned BioBERT classifier achieved the highest macro F1 (0.82), while a lightweight SVM model with TF-IDF features remained competitive (0.78), highlighting their utility in low-resource settings. AINA enables reproducible, large-scale identification of intrinsically bactericidal inorganic NPs. By reducing noise from non-intrinsic contexts, this work provides a foundation for mechanism-aware screening, database construction, and predictive modeling in antimicrobial NP research.

Understanding user intentions is challenging for online platforms. Recent work on intention knowledge graphs addresses this but often lacks focus on connecting intentions, which is crucial for modeling user behavior and predicting future actions. This paper introduces a framework to automatically generate an intention knowledge graph, capturing connections between user intentions. Using the Amazon m2 dataset, we construct an intention graph with 351 million edges, demonstrating high plausibility and acceptance. Our model effectively predicts new session intentions and enhances product recommendations, outperforming previous state-of-the-art methods and showcasing the approach’s practical utility.

Large language models achieve strong performance on many tasks, but their training makes it hard to see which properties of the input support efficient linguistic rule learning.We ask how three cognitively-inspired principles of input design support sample-efficient linguistic rule induction: analogical structure, contrastive learning, and minimal contextual cue. We also ask how their effects compare to those of LLMs on the same controlled tasks.We implement these principles in structured sentence completion tasks that test English verb alternations.Lightweight models trained on hundreds to one-thousand such examples learn the alternation rules with high F1 on these tasks. Ablation studies show that analogical organisation is the main driver of sample efficiency, and contrastive distractors and minimal context help further gains. We also evaluate zero- and few-shot LLMs on the same tasks. In this controlled setting, the lightweight models reach higher F1 with far fewer task-specific data.We treat this contrast as a comparison between learning regimes rather than a general verdict on LLMs.Our results show that careful input organisation supports sample-efficient learning of linguistic rules and reveals distinct learning signatures for trained lightweight models and prompted LLMs.

In this paper, we present the first cross-lingual dataset that captures a transnational cultural phenomenon, focusing on the Chinese and Japanese "Jirai" subculture and its association with risky health behaviors. Our dataset of more than 15,000 annotated social media posts forms the core of JiraiBench, a benchmark designed to evaluate LLMs on culturally specific content. This unique resource allowed us to uncover an unexpected cross-cultural transfer in which Japanese prompts better handle Chinese content, indicating that cultural context can be more influential than linguistic similarity. Further evidence suggests potential cross-lingual knowledge transfer in fine-tuned models. This work proves the indispensable role of developing culturally informed, cross-lingual datasets for creating effective content moderation tools that can protect vulnerable communities across linguistic borders.

Text Generation has achieved remarkable performance using large language models. It has also been recently well-studied that these large language models are capable of creative generation tasks but prominently for high-resource languages. This prompts a fundamental question: Is there a way to utilize these (large) language models for structured poetry generation in a low-resource language, such as Sanskrit? We present Chandomitra, an English input to structured Sanskrit Poetry translation dataset, specifically adhering to the Anushtubh meter. We benchmark various open and closed models, and scrutinize specialized techniques such as constrained decoding and instruction fine-tuning, for the proposed task. Our constrained decoding methodology achieves 99.86% syntactic accuracy in generating metrically valid Sanskrit poetry, outperforming GPT-4o (1-shot: 31.24%). Our best-performing instruction-tuned model, on the other hand, performs better in semantic coherence with the English input, at the expense of slightly lower syntactic accuracy. Human evaluation further reveals that instruction fine-tuned model is better able to capture the poetic aspects.

Large language models (LLMs) show promise in offering emotional support and generating empathetic responses for individuals in distress, but their ability to deliver culturally sensitive support remains underexplored due to a lack of resources. In this work, we introduce , the first dataset designed for this task, spanning four cultures and including 1,729 distress messages, 1,523 cultural signals, and 1,041 support strategies with fine-grained emotional and cultural annotations. Leveraging , we (i) develop and test four adaptation strategies for guiding three state-of-the-art LLMs toward culturally sensitive responses; (ii) conduct comprehensive evaluations using LLM-as-a-Judge, in-culture human annotators, and clinical psychologists; (iii) show that adapted LLMs outperform anonymous online peer responses, and that simple cultural role-play is insufficient for cultural sensitivity; and (iv) explore the application of LLMs in clinical training, where experts highlight their potential in fostering cultural competence in novice therapists.

Open-world Question Answering (OW-QA) over knowledge graphs (KGs) aims to answer questions over incomplete or evolving KGs. Traditional KGQA assumes a closed world where answers must exist in the KG, limiting real-world applicability. In contrast, open- world QA requires inferring missing knowledge based on graph structure and context. Large language models (LLMs) excel at language understanding but lack structured reasoning. Graph neural networks (GNNs) model graph topology but struggle with semantic interpretation. Existing systems integrate LLMs with GNNs or graph retrievers. Some support open-world QA but rely on structural embeddings without semantic grounding. Most assume observed paths or complete graphs, making them unreliable under missing links or multi-hop reasoning. We present GLOW, a hybrid system that combines a pre-trained GNN and an LLM for open-world KGQA. The GNN predicts top-k candidate answers from the graph structure. These, along with relevant KG facts, are serialized into a structured prompt (e.g., triples and candidates) to guide the LLM’s reasoning. This enables joint reasoning over symbolic and semantic signals, without relying on retrieval or fine-tuning. To evaluate generalization, we introduce GLOW-BENCH, a 1,000-question benchmark over incomplete KGs across diverse domains. GLOW outperforms existing LLM–GNN systems on standard benchmarks and GLOW-BENCH, achieving up to 53.3% and an average 38% improvement.

As Large Language Models (LLMs) become increasingly integrated into everyday life and information ecosystems, concerns about their implicit biases continue to persist. While prior work has primarily examined socio-demographic and left–right political dimensions, little attention has been paid to how LLMs align with broader geopolitical value systems, particularly the democracy–authoritarianism spectrum. In this paper, we propose a novel methodology to assess such alignment, combining (1) the F-scale, a psychometric tool for measuring authoritarian tendencies, (2) FavScore, a newly introduced metric for evaluating model favorability toward world leaders, and (3) role-model probing to assess which figures are cited as general role models by LLMs. We find that LLMs generally favor democratic values and leaders, but exhibit increased favorability toward authoritarian figures when prompted in Mandarin. Further, models are found to often cite authoritarian figures as role models, even outside explicitly political contexts. These results shed light on ways LLMs may reflect and potentially reinforce global political ideologies, highlighting the importance of evaluating bias beyond conventional socio-political axes.

Automated feature engineering (AutoFE) liberates data scientists from the burden of manual feature construction. The semantic information of datasets contains rich context information for feature engineering but has been underutilized in many existing AutoFE works. We present PromptFE, a novel AutoFE framework that leverages large language models (LLMs) to automatically construct features in a compact string format and generate semantic explanations based on dataset descriptions. By learning the performance of constructed features in context, the LLM iteratively improves feature construction. We demonstrate through experiments on real-world datasets the superior performance of PromptFE over state-of-the-art AutoFE methods. We verify the impact of dataset semantic information and provide comprehensive study on the LLM-based feature construction process.

Large language models (LLMs) often respond confidently to questions even when they lack the necessary information, leading to hallucinated answers. In this work, we study the problem of (un)answerability detection, focusing on extractive question answering (QA) where the model should determine if a passage contains sufficient information to answer a given question. We propose a simple approach for identifying a direction in the model’s activation space that captures unanswerability and uses it for classification. This direction is selected by applying activation additions during inference and measuring their impact on the model’s abstention behavior. We show that projecting hidden activations onto this direction yields a reliable score for (un)answerability classification. Experiments on two open-weight LLMs and four extractive QA benchmarks show that our method effectively detects unanswerable questions and generalizes better across datasets than existing prompt-based and classifier-based approaches. Moreover, the obtained directions extend beyond extractive QA to unanswerability that stems from factors, such as lack of scientific consensus and subjectivity. Last, causal interventions show that adding or ablating the directions effectively controls the abstention behavior of the model.

Recent advances in reinforcement learning with verifiable rewards (RLVR) show that large language models enhance their reasoning abilities when trained with verifiable signals. However, due to reward sparsity, effectiveness depends heavily on selecting samples of appropriate difficulty. In this work, we present a formal analysis of online difficulty-aware filtering and establish its theoretical foundations. We show that expected policy improvement is lower-bounded by the variance of task-level success probabilities, implying that selecting tasks of intermediate difficulty maximizes learning efficiency. Building on this, we demonstrate that balanced filtering maximizes this lower bound, leading to superior performance and sample efficiency. Evaluations across multiple math reasoning benchmarks validate that balanced filtering consistently enhances convergence speed and final performance, achieving up to +12% gains in less than half the training steps of standard GRPO. By extending our analysis to various reward distributions, we provide a principled foundation for future RLVR curriculum strategies, confirmed through both theoretical analysis and extensive empirical results.

Attention regularisation aims to supervise the attention patterns in language models like BERT. Various studies have shown that using human-annotated rationales, in the form of highlights that explain why a text has a specific label, can have positive effects on model generalisability. In this work, we ask to what extent attention regularisation with human-annotated rationales improve model performance and model robustness, as well as susceptibility to spurious correlations. We compare regularisation on human rationales with randomly selected tokens, a baseline which has hitherto remained unexplored.Our results suggest that often, attention regularisation with randomly selected tokens yields similar improvements to attention regularisation with human-annotated rationales. Nevertheless, we find that human-annotated rationales surpass randomly selected tokens when it comes to reducing model sensitivity to strong spurious correlations.

Categorization is a core component of human linguistic competence. We investigate how a transformer-based language model (LM) learns linguistic categories by comparing its behaviour over the course of training to behaviours which characterize abstract feature–based and concrete exemplar–based accounts of human language acquisition. We investigate how lexical semantic and syntactic categories emerge using novel divergence-based metrics that track learning trajectories using next-token distributions. In experiments with GPT-2 small, we find that (i) when a construction is learned, abstract class-level behaviour is evident at earlier steps than lexical item–specific behaviour, and (ii) that different linguistic behaviours emerge abruptly in sequence at different points in training, revealing that abstraction plays a key role in how LMs learn. This result informs the models of human language acquisition that LMs may serve as an existence proof for.

Social media platforms face escalating challenges in detecting harmful content that promotes muscle dysmorphic behaviors and cognitions (bigorexia). This content can evade moderation by camouflaging as legitimate fitness advice and disproportionately affects adolescent males. We address this challenge with BigTokDetect, a clinically informed framework for identifying pro-bigorexia content on TikTok. We introduce BigTok, the first expert-annotated multimodal benchmark dataset of over 2,200 TikTok videos labeled by clinical psychiatrists across five categories and eighteen fine-grained subcategories. Comprehensive evaluation of state-of-the-art vision-language models reveals that while commercial zero-shot models achieve the highest accuracy on broad primary categories, supervised fine-tuning enables smaller open-source models to perform better on fine-grained subcategory detection. Ablation studies show that multimodal fusion improves performance by 5 to 15 percent, with video features providing the most discriminative signals. These findings support a grounded moderation approach that automates detection of explicit harms while flagging ambiguous content for human review, and they establish a scalable framework for harm mitigation in emerging mental health domains.

While large language models (LLMs) are generally considered proficient in generating language, how similar their language usage is to that of humans remains understudied. In this paper, we test whether models exhibit linguistic convergence, a core pragmatic element of human language communication: do models adapt, or converge, to the linguistic patterns of their user? To answer this, we systematically compare model completions of existing dialogues to original human responses across sixteen language models, three dialogue corpora, and various stylometric features. We find that models strongly converge to the conversation’s style, often significantly overfitting relative to the human baseline. While convergence patterns are often feature-specific, we observe consistent shifts in convergence across modeling settings, with instruction-tuned and larger models converging less than their pretrained and smaller counterparts. Given the differences in human and model convergence patterns, we hypothesize that the underlying mechanisms driving these behaviors are very different.

We expose a critical limitation in current approaches to machine unlearning in language models: despite the apparent success of unlearning algorithms, information about the forgotten data remains linearly decodable from internal representations. To systematically assess this discrepancy, we introduce an interpretable, information-theoretic framework for auditing unlearning using Partial Information Decomposition (PID). By comparing model representations before and after unlearning, we decompose the mutual information with the forgotten data into distinct components, formalizing the notions of unlearned and residual knowledge. Our analysis reveals that redundant information, shared across both models, constitutes residual knowledge that persists post-unlearning and correlates with susceptibility to known adversarial reconstruction attacks. Leveraging these insights, we propose a representation-based risk score that can guide abstention on sensitive inputs at inference time, providing a practical mechanism to mitigate privacy leakage. Our work introduces a principled, representation-level audit for unlearning, offering theoretical insight and actionable tools for safer deployment of language models.

We consider coverless steganography where a Large Language Model (LLM) is used to generate stego-texts in combination with arithmeticic coding. An efficient method should embed secret bits in as few language tokens as possible while keeping the stego-text as natural as possible. We show that this problem is equivalent to maximizing the entropy of a replacement probability distribution of the next token generation, subject to a constraint on the divergence between the new distribution and the original one produced by the LLM. A closed-form solution is provided under either the KL divergence or the total variation constraint. Several important practical issues are also tackled: 1) An often-overlooked tokenization mismatch issue is resolved with a simple prompt selection approach, 2) The combination of the optimized distribution and the vocabulary truncation technique is considered, and 3) The incorporation of the proposed approach with existing (potentially non arithemtic coding based) techniques, e.g., the Discop technique.

Continual learning in natural language processing plays a crucial role in adapting to evolving data and preventing catastrophic forgetting. Despite significant progress, existing methods still face challenges, such as inefficient parameter reuse across tasks, risking catastrophic forgetting when tasks are dissimilar, and the unnecessary introduction of new parameters for each task, which hampers knowledge sharing among similar tasks. To tackle these issues, we propose a Sparse Adapter Fusion Method (SAFM), which dynamically fuses old and new adapters to address these challenges. SAFM operates in two stages: the decision stage and the tuning stage. In the decision stage, SAFM determines whether to incorporate a new adapter, reuse an existing one, or add an empty adapter. The architecture search procedure, designed to prioritize reusing or adding empty adapters, minimizes parameter consumption and maximizes reuse. In the tuning stage, SAFM especially facilitates a layer-wise loss to encourage differentiation between adapters, effectively capturing knowledge within the same task. Experimental results consistently show that SAFM outperforms state-of-the-art (SOTA) methods, achieving comparable performance while utilizing less than 60% of the parameters.

Prompt optimization (PO) provides a practical way to improve response quality when users lack the time or expertise to manually craft effective prompts. Existing methods typically rely on LLMs’ self-generation ability to optimize prompts. However, due to limited downward compatibility, the instruction-heavy prompts generated by advanced LLMs can overwhelm lightweight inference models and degrade response quality, while also lacking interpretability due to implicit optimization. In this work, we rethink prompt optimization through the lens of explicit and interpretable design. We first identify a set of model-agnostic prompt quality merits and empirically validate their effectiveness in enhancing prompt and response quality. We then introduce MePO, a merit-guided, locally deployable prompt optimizer trained on our merit-guided prompt preference dataset generated by a lightweight LLM. MePO avoids online optimization, reduces privacy concerns, and, by learning clear, interpretable merits, generalizes effectively to both large-scale and lightweight inference models. Experiments demonstrate that MePO achieves better results across diverse tasks and model types, offering a scalable and robust solution for real-world deployment. The code, model and dataset can be found in https://github.com/MidiyaZhu/MePO.

The increasing prevalence of mental disorders globally highlights the urgent need for effective digital screening methods that can be used in multilingual contexts. Most existing studies, however, focus on English data, overlooking critical mental health signals that may be present in non-English texts. To address this gap, we present a survey of the detection of mental disorders using social media data beyond the English language. We compile a comprehensive list of 108 datasets spanning 25 languages that can be used for developing NLP models for mental health screening. In addition, we discuss the cultural nuances that influence online language patterns and self-disclosure behaviors, and how these factors can impact the performance of NLP tools. Our survey highlights major challenges, including the scarcity of resources for low- and mid-resource languages and the dominance of depression-focused data over other disorders. By identifying these gaps, we advocate for interdisciplinary collaborations and the development of multilingual benchmarks to enhance mental health screening worldwide.

Large Language Models (LLMs) have demonstrated remarkable capabilities across a wide range of instruction-following tasks, yet their grasp of nuanced social science concepts remains underexplored. This paper examines whether LLMs can identify and classify fine-grained forms of populism, a complex and contested concept in both academic and media debates. To this end, we curate and release novel datasets specifically designed to capture populist discourse. We evaluate a range of pre-trained (large) language models, both open-weight and proprietary, across multiple prompting paradigms. Our analysis reveals notable variation in performance, highlighting the limitations of LLMs in detecting populist discourse. We find that a fine-tuned RoBERTa classifier vastly outperforms all new-era instruction-tuned LLMs, unless fine-tuned. Additionally, we apply our best-performing model to analyze campaign speeches by Donald Trump, extracting valuable insights into his strategic use of populist rhetoric. Finally, we assess the generalizability of these models by benchmarking them on campaign speeches by European politicians, offering a lens into cross-context transferability in political discourse analysis. In this setting, we find that instruction-tuned LLMs exhibit greater robustness on out-of-domain data.

Generative methods have recently gained traction in biological and chemical named entity recognition for their ability to overcome tagging limitations and better capture entity-rich contexts. However, under a few-shot environment, they struggle with the scarcity of annotated data and the structural complexity of biological and chemical entities—particularly nested and discontinuous ones—leading to incorrect recognition and error propagation during generation. To address these challenges, we propose SCoNE, a Self-Correcting and Noise-Augmented Method for Complex Biological and Chemical Named Entity Recognition. Specifically, we introduce a Noise Augmentation Module to enhance training diversity and guide the model to better learn complex entity structures. Besides, we design a Confidence-based Self-Correction Module that identifies low-confidence outputs and revises them to improve generation robustness. Benefiting from these designs, our method outperforms the baselines by 1.80 and 2.73 F1-score on the CHEMDNER and microbial ecology dataset Florilege, highlighting its effectiveness in biological and chemical named entity recognition.

The present benchmarks for testing the audio modality of multimodal large language models concentrate on testing various audio tasks such as speaker diarization or gender identification in isolation. Whether a multimodal model can answer the questions that require reasoning skills to combine audio tasks of different categories cannot be verified with their use. To address this issue, we propose Audio Reasoning Tasks (ART), a new benchmark for assessing the ability of multimodal models to solve problems that require reasoning over audio signal.

Prior work has successfully applied Reinforcement Learning (RL) to mathematical reasoning—where rules and correctness are well-defined. Yet, generalizing these methods to broader reasoning domains remains challenging due to limited data and the lack of verifiable rewards for unstructured domains. In this work, we propose NEMOTRON-CROSSTHINK, a framework that systematically incorporates multi-domain corpora into RL training to improve generalization across diverse reasoning tasks. NEMOTRON-CROSSTHINK addresses key challenges by (1) combining data from varied sources; (2) applying structured templates to control answer-space complexity; (3) filtering for verifiable answers; and (4) optimizing data blending strategies to utilize multi-source data effectively. This enables scalable and verifiable reward modeling beyond math and demonstrates improved accuracies on both math (MATH-500: +30.1%, AMC23: +27.5%) and non-math reasoning benchmarks (MMLU-PRO: +12.8%, GPQA-DIAMOND: +11.3%, AGIEVAL: +15.1%, SUPERGPQA: +3.8%). Moreover, NEMOTRON-CROSSTHINK exhibits significantly improved response efficiency—using 28% fewer tokens for correct answers—highlighting more focused and effective reasoning. Through NEMOTRON-CROSSTHINK, we demonstrate that integrating multi-domain, multi-format data in RL leads to more accurate, efficient, and generalizable LLMs. All of our datasets are available on HuggingFace.

As Large Language Models (LLMs) continue to evolve, ensuring their safety across multiple languages has become a critical concern. While LLMs demonstrate impressive capabilities in English, their safety mechanisms may not generalize effectively to other languages, leading to disparities in toxicity detection, bias mitigation, and harm prevention. This systematic review examines the multilingual safety of LLMs by synthesizing findings from recent studies that evaluate their robustness across diverse linguistic and cultural contexts beyond English language. Our review explores the methodologies used to assess multilingual safety, identifies challenges such as dataset availability and evaluation biases. Based on our analysis we highlight gaps in multilingual safety research and provide recommendations for future work. This review aims to contribute to the development of fair and effective safety mechanisms for LLMs across all languages. We provide the extracted data in an interactive Streamlit dashboard, enabling transparent access to the raw data and allowing for continuous updates.

Graphical User Interface (GUI) Agents, powered by multimodal large language models (MLLMs), have shown great potential for task automation on computing devices such as computers and mobile phones. However, existing agents face challenges in multi-step reasoning and reliance on textual annotations, limiting their effectiveness. We introduce InfiGUIAgent, an MLLM-based GUI Agent trained with a two-stage supervised fine-tuning pipeline. Stage 1 enhances fundamental skills such as GUI understanding and grounding, while Stage 2 integrates hierarchical reasoning and expectation-reflection reasoning skills using synthesized data to enable native reasoning abilities of the agents. InfiGUIAgent achieves competitive performance on several GUI benchmarks, highlighting the impact of native reasoning skills in enhancing GUI interaction for automation tasks.

We introduce Cetvel, a comprehensive benchmark designed to evaluate large language models (LLMs) in Turkish. Existing Turkish benchmarks often lack either task diversity or culturally relevant content, or both. Cetvel addresses these gaps by combining a broad range of both discriminative and generative tasks ensuring content that reflects the linguistic and cultural richness of Turkish language. Cetvel covers 23 tasks grouped into seven categories, including tasks such as grammatical error correction, machine translation, and question answering rooted in Turkish history and idiomatic language. We evaluate 33 open-weight LLMs (up to 70B parameters) covering different model families and instruction paradigms. Our experiments reveal that Turkish-centric instruction-tuned models generally underperform relative to multilingual or general-purpose models (e.g. Llama 3 and Mistral), despite being tailored for the language. Moreover, we show that tasks such as grammatical error correction and extractive question answering are particularly discriminative in differentiating model capabilities. Cetvel offers a comprehensive and culturally grounded evaluation suite for advancing the development and assessment of LLMs in Turkish.

Lexical semantics is concerned with both the multiple senses a word can adopt in different contexts, and the semantic relations that exist between meanings of different words. To investigate them, Contextualized Language Models are a valuable tool that provides context-sensitive representations that can be used to investigate lexical meaning. Recent works like XL-LEXEME have leveraged the task of Word-in-Context to fine-tune them to get more semantically accurate representations, but Word-in-Context only compares occurrences of the same lemma, limiting the range of captured information. In this paper, we propose an extension, Concept Differentiation, to include inter-words scenarios. We provide a dataset for this task, derived from SemCor data. Then we fine-tune several representation models on this dataset. We call these models Concept-Aligned Embeddings (CALE). By challenging our models and other models on various lexical semantic tasks, we demonstrate that the proposed models provide efficient multi-purpose representations of lexical meaning that reach best performances in our experiments. We also show that CALE’s fine-tuning brings valuable changes to the spatial organization of embeddings.

Evaluating attribute control success in controllable text generation and related generation tasks typically relies on pretrained classifiers. We show that this widely used classify-and-count approach yields biased and inconsistent results, with estimates varying significantly across classifiers. We frame control success estimation as a quantification task and apply a hybrid Bayesian method that combines classifier predictions with a small number of human labels for calibration. To test our approach, we collected a two-modality test dataset consisting of 600 human-rated samples and 60,000 automatically rated samples. Our experiments show that our approach produces robust estimates of control success across both text and text-to-image generation tasks, offering a principled alternative to current evaluation practices.

Membership inference attacks (MIAs) aim to determine whether a specific example was used to train a given language model. While prior work has explored prompt-based attacks such as ReCALL, these methods rely heavily on the assumption that using known non-members as prompts reliably suppresses the model’s responses to non-member queries. We propose EM-MIA, a new membership inference approach that iteratively refines prefix effectiveness and membership scores using an expectation-maximization strategy without requiring labeled non-member examples. To support controlled evaluation, we introduce OLMoMIA, a benchmark that enables analysis of MIA robustness under systematically varied distributional overlap and difficulty. Experiments on WikiMIA and OLMoMIA show that EM-MIA outperforms existing baselines, particularly in settings with clear distributional separability. We highlight scenarios where EM-MIA succeeds in practical settings with partial distributional overlap, while failure cases expose fundamental limitations of current MIA methods under near-identical conditions. We release our code and evaluation pipeline to encourage reproducible and robust MIA research.

STEM Mental models can play a critical role in assessing students’ conceptual understanding of a topic. They not only offer insights into what students know but also into how effectively they can apply, relate to, and integrate concepts across various contexts. Thus, students’ responses are critical markers of the quality of their understanding and not entities that should be merely graded. However, inferring these mental models from student answers is challenging as it requires deep reasoning skills. We propose MMGrader, an approach that infers the quality of students’ mental models from their multimodal responses using concept graphs as an analytical framework. In our evaluation with 9 openly available models, we found that the best-performing models fall short of human-level performance. This is because they only achieved an accuracy of approximately 40%, a prediction error of 1.1 units, and a scoring distribution fairly aligned with human scoring patterns. With improved accuracy, these can be highly effective assistants to teachers in inferring the mental models of their entire classrooms, enabling them to do so efficiently and help improve their pedagogies more effectively by designing targeted help sessions and lectures that strengthen areas where students collectively demonstrate lower proficiency.

The ability of Large Language Models (LLMs) to mimic human behavior triggered a plethora of computational social science research, assuming that empirical studies of humans can be conducted with AI agents instead. Since there have been conflicting research findings on whether and when this hypothesis holds, there is a need to better understand the differences in their experimental designs. We focus on replicating the behavior of social network users with the use of LLMs for the analysis of communication on social networks. First, we provide a formal framework for the simulation of social networks, before focusing on the sub-task of imitating user communication. We empirically test different approaches to imitate user behavior on in English and German. Our findings suggest that social simulations should be validated by their empirical realism measured in the setting in which the simulation components were fitted. With this paper, we argue for more rigor when applying generative-agent-based modeling for social simulation.

For socially sensitive tasks like hate speech detection, the quality of explanations from Large Language Models (LLMs) is crucial for factors like user trust and model alignment. While Persona prompting (PP) is increasingly used as a way to steer model towards user-specific generation, its effect on model rationales remains underexplored. We investigate how LLM-generated rationales vary when conditioned on different simulated demographic personas. Using datasets annotated with word-level rationales, we measure agreement with human annotations from different demographic groups, and assess the impact of PP on model bias and human alignment. Our evaluation across three LLMs results reveals three key findings: (1) PP improving classification on the most subjective task (hate speech) but degrading rationale quality. (2) Simulated personas fail to align with their real-world demographic counterparts, and high inter-persona agreement shows models are resistant to significant steering. (3) Models exhibit consistent demographic biases and a strong tendency to over-flag content as harmful, regardless of PP. Our findings reveal a critical trade-off: while PP can improve classification in socially-sensitive tasks, it often comes at the cost of rationale quality and fails to mitigate underlying biases, urging caution in its application.

Detecting hyperpartisan narratives and Population Replacement Conspiracy Theories (PRCT) is essential to addressing the spread of misinformation. These complex narratives pose a significant threat, as hyperpartisanship drives political polarisation and institutional distrust, while PRCTs directly motivate real-world extremist violence, making their identification critical for social cohesion and public safety. However, existing resources are scarce, predominantly English-centric, and often analyse hyperpartisanship, stance, and rhetorical bias in isolation rather than as interrelated aspects of political discourse. To bridge this gap, we introduce PartisanLens, the first multilingual dataset of 1617 hyperpartisan news headlines in Spanish, Italian, and Portuguese, annotated in multiple political discourse aspects. We first evaluate the classification performance of widely used Large Language Models (LLMs) on this dataset, establishing robust baselines for the classification of hyperpartisan and PRCT narratives. In addition, we assess the viability of using LLMs as automatic annotators for this task, analysing their ability to approximate human annotation. Results highlight both their potential and current limitations. Next, moving beyond standard judgments, we explore whether LLMs can emulate human annotation patterns by conditioning them on socio-economic and ideological profiles that simulate annotator perspectives. At last, we provide our resources and evaluation; PartisanLens supports future research on detecting partisan and conspiratorial narratives in European contexts.

Neuro-symbolic NLP methods aim to leverage the complementary strengths of large language models and formal logical solvers. However, current approaches are mostly static in nature, i.e., the integration of a target solver is predetermined at design time, hindering the ability to employ diverse formal inference strategies. To address this, we introduce an adaptive, multi-paradigm, neuro-symbolic inference framework that: (1) automatically identifies formal reasoning strategies from problems expressed in natural language; and (2) dynamically selects and applies specialized formal logical solvers via autoformalization interfaces. Extensive experiments on individual and multi-paradigm reasoning tasks support the following conclusions: LLMs are effective at predicting the necessary formal reasoning strategies with an accuracy above 90%. This enables flexible integration with formal logical solvers, resulting in our framework outperforming competing baselines – by 17% and 6% compared to GPT-4o and DeepSeek-V3.1, respectively. Moreover, adaptive reasoning can even positively impact pure LLM methods, yielding gains of 10%, 5%, and 6% on zero-shot, CoT, and CoT_sym settings with GPT-4o. Finally, although smaller models struggle with adaptive neuro-symbolic reasoning, post-training offers a viable path to improvement. Overall, this work establishes the foundations for adaptive LLM-symbolic reasoning, offering a path forward for unifying material and formal inferences on heterogeneous reasoning challenges.The code and data are available at https://github.com/idiap/adaptive_symbolic_reasoning.

Large Language Models (LLMs) excel in numerous and varied tasks. Yet, the mechanisms that underlie this success remain insufficiently understood. In particular, the size and the limited transparency of their pre-training materials make it difficult to state what the properties of the pre-training material are when compared to the test data. In this paper, we investigate whether LLMs learned generalized linguistic abstraction or rely on surface-level features, like lexical patterns, that match their pre-training data. We explore this by examining the relationship between lexical overlap of test data and task performance. We observe that lexical overlap with the pre-training material is mostly beneficial to model performance on tasks requiring functional linguistic knowledge. To further explore the impact of lexical features, we also demonstrate that LLMs are fragile with respect to lexical perturbations that preserve semantics. While we expected models to rely on lexical overlap between test instances and pre-training data for tasks requiring functional knowledge, lexical perturbations reveal that models also exhibit, to a lesser extent, this dependence for tasks requiring formal linguistic knowledge.

Large language models (LLMs) are increasingly applied to tasks involving causal reasoning. However, current benchmarks often rely on string matching or surface-level metrics that fail to assess whether a model’s output is formally valid under causal semantics. We propose DoVerifier, a symbolic verification framework that checks whether LLM-generated causal expressions are derivable from a given causal graph using rules from do-calculus and probability theory. This allows us to recover correct answers that would otherwise be marked incorrect due to superficial differences. Evaluations on synthetic data and causal QA benchmarks show that DoVerifier more accurately captures semantic correctness than standard metrics, offering a more rigorous and informative way to evaluate LLMs on causal tasks.

Game-theoretic interactions between agents with large language models (LLMs) have revealed many emergent capabilities, yet the linguistic diversity of these interactions has not been sufficiently quantified. In this paper, we present the Conversational Robustness Evaluation Score: CORE, a metric to quantify the effectiveness of language use within multi-agent systems across different game-theoretic interactions. CORE integrates measures of cluster entropy, lexical repetition, and semantic similarity, providing a direct lens of dialog quality. We apply CORE to pairwise LLM dialogs across competitive, cooperative, and neutral settings, further grounding our analysis in Zipf’s and Heaps’ Laws to characterize word frequency distributions and vocabulary growth. Our findings show that cooperative settings exhibit both steeper Zipf distributions and higher Heap exponents, indicating more repetition alongside greater vocabulary expansion. In contrast, competitive interactions display lower Zipf and Heaps exponents, reflecting less repetition and more constrained vocabularies. These results provide new insights into how social incentives influence language adaptation, and highlight CORE as a robust diagnostic for measuring linguistic robustness in multi-agent LLM systems.

Scaling data quantity is essential for large language models (LLMs), yet recent findings show that data quality can significantly boost performance and training efficiency. We introduce a German-language dataset curation pipeline that combines heuristic and model-based filtering techniques with synthetic data generation. We use our pipeline to create Aleph-Alpha-GermanWeb, a large-scale German pre-training dataset which draws from: (1) Common Crawl web data, (2) FineWeb2, and (3) synthetically-generated data conditioned on actual, organic web data. We evaluate our dataset by pre-training both a 1B Llama-style model and an 8B tokeniser-free hierarchical autoregressive transformer (HAT). A comparison on German-language benchmarks, including MMMLU, shows significant performance gains of Aleph-Alpha-GermanWeb over FineWeb2 alone. This advantage holds at the 8B scale even when FineWeb2 is enriched by human-curated high-quality data sources such as Wikipedia. Our findings support the growing body of evidence that model-based data curation and synthetic data generation can significantly enhance LLM pre-training datasets.

Large language models achieve state-of-the-art performance but are increasingly costly to fine-tune. Prompt tuning is a parameter-efficient fine-tuning method that addresses parameter-efficiency by learning prompt embeddings, but these embeddings are typically tied to the model’s hidden dimensionality, limiting parameter saving. In this paper, we propose Ultra-Low-dimensional Prompt Tuning (ULPT), a simple yet effective method that optimizes prompts in a low-dimensional space (e.g., 2D) and uses a frozen random matrix for up-projection. ULPT can achieve 98% reduction in the training parameters compared to vanilla prompt tuning while preserving performance. Our extensive experiments across over 20 NLP tasks demonstrate that ULPT consistently outperforms recent parameter-efficient tuning methods using significantly fewer parameters, making it well-suited as a storage-efficient framework for massive LLM customization.

We model Semantic Self-Verification (SSV) as the problem of determining whether a statement accurately characterizes its own semantic properties within a given interpretive framework that formalizes a challenge in AI safety and fairness: can an AI system verify that it has correctly interpreted rules intended to govern its behavior? We prove that SSV, in this specification, is NP-complete by constructing a polynomial-time reduction from 3-Satisfiability (3-SAT). Our reduction maps a 3-SAT formula to an instance of SSV involving ambiguous terms with binary interpretations and semantic constraints derived from logical clauses. This establishes that even simplified forms of semantic self-verification should face computational barriers. The NP-complete lower bound has implications for AI safety and fairness approaches that rely on semantic interpretation of instructions, including but not limited to constitutional AI, alignment via natural language, and instruction-following systems. Any approach where an AI system verify its understanding of directives faces this computational barrier. We argue that more realistic verification scenarios likely face even greater complexity.

We present STAMP (Selective Task-Aware Mechanism for Text Privacy), a new framework for task-aware text privatization that achieves an improved privacy–utility trade-off. STAMP selectively allocates privacy budgets across tokens by jointly considering (i) each token’s importance to the downstream task (as measured via a task- or query-specific representation), and (ii) its privacy sensitivity (e.g., names, dates, identifiers). This token-level partitioning enables fine-grained, group-wise control over the level of noise applied to different parts of the input, balancing privacy protection with task relevance. To privatize individual token embeddings, we introduce the polar mechanism, which perturbs only the direction of embeddings on the unit sphere while preserving their magnitude. Decoding is performed via cosine nearest-neighbor search, aligning the perturbation geometry with the decoding geometry. Unlike isotropic noise mechanisms, the polar mechanism maintains semantic neighborhoods in the embedding space and better preserves downstream utility. Experimental evaluations on SQuAD, Yelp, and AG News datasets demonstrate that STAMP, when combined with the normalized polar mechanism, consistently achieves superior privacy–utility trade-offs across varying per-token privacy budgets.

Reliably ensuring Large Language Models (LLMs) follow complex instructions is a critical challenge, as existing benchmarks often fail to reflect real-world use or isolate compliance from task success. We introduce MOSAIC (MOdular Synthetic Assessment of Instruction Compliance), a modular framework that uses a dynamically generated dataset with up to 20 application-oriented generation constraints to enable a granular and independent analysis of this capability.Our evaluation of five LLMs from different families based on this new benchmark demonstrates that compliance is not a monolithic capability but varies significantly with constraint type, quantity, and position. The analysis reveals model-specific weaknesses, uncovers synergistic and conflicting interactions between instructions, and identifies distinct positional biases such as primacy and recency effects. These granular insights are critical for diagnosing model failures and developing more reliable LLMs for systems that demand strict adherence to complex instructions.

As Large Language Models(LLMs) increasingly power chatbots, social media, and other interactive platforms, the ability to detect AI in conversational settings is critical for ensuring transparency and preventing potential misuse. However, existing detection methods focus on static, document-level content, overlooking the dynamic nature of dialogues. To address this, we propose an utterance-level detection framework, which integrates features from individual and combined analysis of dialogue participants’ responses to detect LLM-generated text under conversational setting. Leveraging a transformer-based recurrent architecture and a curated dataset of human-human, human-LLM, and LLM-LLM dialogues, this framework achieves an accuracy of 98.14% with high inference speed, supported by extensive results of experiments on different models and settings. This work provides an effective solution for detecting LLM-generated text in real-time conversations, promoting transparency, and mitigating risks of misuse.

Real-world clinical text-to-SQL requires reasoning over heterogeneous EHR tables, temporal windows, and patient-similarity cohorts to produce executable queries. We introduce ClinSQL, a benchmark of 633 expert-annotated tasks on MIMIC-IV v3.1 that demands multi-table joins, clinically meaningful filters, and executable SQL. Solving ClinSQL entails navigating schema metadata and clinical coding systems, handling long contexts, and composing multi-step queries beyond traditional text-to-SQL. We evaluate 20 proprietary and open-source models under Chain-of-Thought self-refinement and use rubric-based SQL analysis with execution checks that prioritize critical clinical requirements. Despite recent advances, performance remains far from clinical reliability: on the test set, GPT-5-mini attains 74.7% execution score, DeepSeek-R1 leads open-source at 69.2% and Gemini-2.5-Pro drops from 85.5% on Easy to 67.2% on Hard. Progress on ClinSQL marks tangible advances toward clinically reliable text-to-SQL for real-world EHR analytics.

We present iBERT (interpretable-BERT), an encoder to produce inherently interpretable and controllable embeddings - designed to modularize and expose the discriminative cues present in language, such as semantic or stylistic structure. Each input token is represented as a sparse, non-negative mixture over k context-independent sense vectors, which can be pooled into sentence embeddings or used directly at the token level. This enables modular control over representation, before any decoding or downstream use.To demonstrate our model’s interpretability, we evaluate it on a suite of style-focused tasks. On the STEL benchmark, it improves style representation effectiveness by ~8 points over SBERT-style baselines, while maintaining competitive performance on authorship verification. Because each embedding is a structured composition of interpretable senses, we highlight how specific style attributes get assigned to specific sense vectors. While our experiments center on style, iBERT is not limited to stylistic modeling. Its structural modularity is designed to interpretably decompose whichever discriminative signals are present in the data — enabling generalization even when supervision blends semantic or stylistic factors.

This paper investigates the real-world vulnerabilities of audio-based large language models (ALLMs), such as Qwen2-Audio. We first demonstrate that an adversary can craft stealthy audio perturbations to manipulate ALLMs into exhibiting specific targeted behaviors, such as eliciting responses to wake-keywords (e.g., "Hey Qwen"), or triggering harmful behaviors (e.g., "Change my calendar event"). Subsequently, we show that playing adversarial background noise during user interaction with the ALLMs can significantly degrade the response quality. Crucially, our research illustrates the scalability of these attacks to real-world scenarios, impacting other innocent users when these adversarial noises are played through the air. Further, we discuss the transferability of the attack and potential defensive measures.

Large language models (LLMs) are highly sensitive to subtle changes in prompt phrasing, posing challenges for reliable auditing. Prior methods often apply unconstrained prompt paraphrasing, which risk missing linguistic and demographic factors that shape authentic user interactions. We introduce AUGMENT (Automated User-Grounded Modeling and Evaluation of Natural Language Transformations), a framework for generating controlled paraphrases, grounded in user behaviors. AUGMENT leverages linguistically informed rules and enforces quality through checks on instruction adherence, semantic similarity, and realism, ensuring paraphrases are both reliable and meaningful for auditing. Through case studies on the BBQ and MMLU datasets, we show that controlled paraphrases uncover systematic weaknesses that remain obscured under unconstrained variation. These results highlight the value of the AUGMENT framework for reliable auditing.

We present AutoBool, a reinforcement learning (RL) framework that trains large language models (LLMs) to generate effective Boolean queries for medical systematic reviews. Boolean queries are the primary mechanism for literature retrieval in this domain and must achieve high recall while maintaining reasonable precision—a challenging balance that existing prompt-based LLM approaches often struggle to achieve.A major limitation in this space is the lack of ground-truth best Boolean queries for each topic, which makes supervised fine-tuning impractical. AutoBool addresses this challenge by leveraging RL to directly optimize query generation against retrieval performance metrics, without requiring ideal target queries. To support this effort, we create and release the largest dataset of its kind: 65 588 topics in total for training and evaluating the task of automatic Boolean query formulation.Experiments on our new dataset and two established datasets (CLEF TAR and Seed Collection) show that AutoBool significantly outperforms zero-shot/few-shot prompting and matches or exceeds the effectiveness of much larger GPT-based models (e.g., GPT-4o, O3) using smaller backbones. It also approaches effectiveness of expert-authored queries while retrieving 10–16 times fewer documents. Ablation studies reveal the critical roles of model backbone, size, decoding temperature, and prompt design. Code and data are available at https://github.com/ielab/AutoBool.

Large language models (LLMs) often generate off-domain or harmful responses when deployed in specialized, high-stakes domains, motivating the need for rigorous LLM domain certification. While the VALID algorithm (Emde et al., 2025) achieves formal domain certificate guarantee using a guide model G trained from scratch on in-domain data, it suffers from poor generalization due to limited training. In this work, we propose PRISM, a novel approach that overcomes this key limitation by leveraging pretrained language models as guide models, enhanced via contrastive fine-tuning to sharply distinguish acceptable from refused content. We explore and experiment variants of PRISM with different loss functions to ensure that the model exploits the rich world knowledge of pretrained models while aligned to the target domain. We show that two variants of PRISM, PRISM-BC and PRISM-GA, achieve superior OOD rejection and tighter certification bounds across eight diverse data regimes and perturbations, establishing a more reliable approach to domain-adherent LLM deployment.

We introduce TDFlow, a novel test-driven agentic workflow that frames repository-scale software engineering as a test-resolution task, specifically designed to solve human-written tests. Given a set of tests, TDFlow repeatedly proposes, revises, and debugs repository-scale patches using precisely engineered sub-agents and tightly constrained tools. The workflow decomposes software engineering program repair into four components governed by respective sub-agents. This simple, forced decoupling of patch proposing, debugging, patch revision, and optional test generation (1) reduces long-context burden on any individual sub-agent, (2) focuses each sub-agent on specific, pre-defined sub-tasks, and (3) allows for specialized performance improvement on specific sub-tasks. When provided human-written tests, TDFlow attains 88.8% pass rate on SWE-Bench Lite (an absolute improvement of 27.8% over the next best baseline) and 94.3% on SWE-Bench Verified. In this work, we further show that the primary obstacle to human-level software engineering performance lies within writing successful reproduction tests. Manual inspection of the 800 TDFlow runs within SWE-Bench Lite and Verified uncover only 7 instances of test hacking, which were subsequently counted as failures. We envision a human-LLM interactive system powered by TDFlow where human developers write tests solved by LLM systems. Together, these results show that modern LLMs, when embedded in a narrowly engineered, test-driven workflow, already achieve human-level test resolution – with the final frontier for fully autonomous repository repair being accurate reproduction test generation.

Understanding narratives requires identifying which events are most salient for a story’s progression. We present a contrastive learning framework for modeling narrative salience that learns story embeddings from narrative twins: stories that share the same plot but differ in surface form. Our model is trained to distinguish a story from both its narrative twin and a distractor with similar surface features but different plot. Using the resulting embeddings, we evaluate four narratologically motivated operations for inferring salience (deletion, shifting, disruption, and summarization). Experiments on short narratives from the ROCStories corpus and longer Wikipedia plot summaries show that contrastively learned story embeddings outperform a masked-language-model baseline, and that summarization is the most reliable operation for identifying salient sentences. If narrative twins are not available, random dropout can be used to generate the twins from a single story. Effective distractors can be obtained either by prompting LLMs or, in long-form narratives, by using different parts of the same story.

Large language models (LLMs) struggle with ex-ante reasoning—making inferences or predictions without access to future information. Even under explicit temporal cutoffs, they often rely on internalized post-cutoff knowledge. To systematically evaluate this issue, we introduce a benchmark that assesses LLMs’ ex-ante inference ability across four tasks: stock prediction, question answering, Wikipedia event generation, and scientific publication generation. We quantify temporal leakage using a leakage rate metric, which measures models’ reliance on future information beyond cutoff timestamps, and a quality measure that evaluates task performance. Experimental results show that LLMs frequently violate temporal constraints across tasks, revealing persistent challenges in ex-ante reasoning. Our benchmark serves as a rigorous testbed for studying temporal reasoning in time-sensitive contexts and provides complete datasets, results, and evaluation resources to support future research on improving temporal consistency in modern LLMs.

Detecting mental health crisis situations such as suicide ideation, rape, domestic violence, child abuse, and sexual harassment is a critical yet underexplored challenge for language models. When such situations arise during user–model interactions, models must reliably flag them, as failure to do so can have serious consequences. In this work, we introduce CRADLE BENCH, a benchmark for multi-faceted crisis detection. Unlike previous efforts that focus on a limited set of crisis types, our benchmark covers seven types defined in line with clinical standards and is the first to incorporate temporal labels. Our benchmark provides 600 clinician-annotated evaluation examples and 420 development examples, together with a training corpus of around 4K examples automatically labeled using a majority-vote ensemble of multiple language models, which significantly outperforms single-model annotation. We further fine-tune six crisis detection models on subsets defined by consensus and unanimous ensemble agreement, providing complementary models trained under different agreement criteria.Content warning: This paper discusses sensitive topics such as suicide ideation, self-harm, rape, domestic violence, and child abuse.

Geocoding is the task of linking a location reference to an actual geographic location and is essential for many downstream analyses of unstructured text. In this paper, we explore the challenging setting of geocoding compositional location references. Building on recent work demonstrating LLMs’ abilities to reason over geospatial data, we evaluate LLMs’ geospatial knowledge versus reasoning skills relevant to our task. Based on these insights, we propose an LLM-based strategy for geocoding compositional location references. We show that our approach improves performance for the task and that a relatively small fine-tuned LLM can achieve comparable performance with much larger off-the-shelf models.

Adapting large language models to full document translation remains challenging due to the difficulty of capturing long-range dependencies and preserving discourse coherence throughout extended texts. While recent agentic machine translation systems mitigate context window constraints through multi-agent orchestration and persistent memory, they require substantial computational resources and are sensitive to memory retrieval strategies. We introduce TransGraph, a discourse-guided framework that explicitly models inter-chunk relationships through structured discourse graphs and selectively conditions each translation segment on relevant graph neighbourhoods rather than relying on sequential or exhaustive context. Across three document-level MT benchmarks spanning six languages and diverse domains, TransGraph consistently surpasses strong baselines in translation quality and terminology consistency while incurring significantly lower token overhead.

Workflows are a fundamental component of automation in enterprise platforms, enabling the orchestration of tasks, data processing, and system integrations. Despite being widely used, building workflows can be complex, often requiring manual configuration through low-code platforms or visual programming tools. To simplify this process, we explore the use of generative foundation models, particularly vision-language models (VLMs), to automatically generate structured workflows from visual inputs. Translating hand-drawn sketches or computer-generated diagrams into executable workflows is challenging due to the ambiguity of free-form drawings, variations in diagram styles, and the difficulty of inferring execution logic from visual elements. To address this, we introduce StarFlow, a framework for generating structured workflow outputs from sketches using vision-language models. We curate a diverse dataset of workflow diagrams – including synthetic, manually annotated, and real-world samples – to enable robust training and evaluation. We finetune and benchmark multiple vision-language models, conducting a series of ablation studies to analyze the strengths and limitations of our approach. Our results show that finetuning significantly enhances structured workflow generation, outperforming large vision-language models on this task.

Ensuring that large language models (LLMs) are both helpful and harmless is a critical challenge, as overly strict constraints can lead to excessive refusals, while permissive models risk generating harmful content. Existing approaches, such as reinforcement learning from human feedback (RLHF) and direct preference optimization (DPO), attempt to balance these trade-offs but suffer from performance conflicts, limited controllability, and poor extendability. To address these issues, we propose Preference Vector, a novel framework inspired by task arithmetic. Instead of optimizing multiple preferences within a single objective, we train separate models on individual preferences, extract behavior shifts as preference vectors, and dynamically merge them at test time. This modular approach enables fine-grained, user-controllable preference adjustments and facilitates seamless integration of new preferences without retraining. Experiments show that our proposed Preference Vector framework improves helpfulness without excessive conservatism, allows smooth control over preference trade-offs, and supports scalable multi-preference alignment.

The miscalibration of Large Reasoning Models (LRMs) undermines their reliability in high-stakes domains, necessitating methods to accurately estimate the confidence of their long-form, multi-step outputs. To address this gap, we introduce the Reasoning Model Confidence estimation Benchmark (RMCB), a public resource of 347,496 reasoning traces from six popular LRMs across different architectural families. The benchmark is constructed from a diverse suite of datasets spanning high-stakes domains, including clinical, financial, legal, and mathematical reasoning, alongside complex general reasoning benchmarks, with correctness annotations provided for all samples. Using RMCB, we conduct a large-scale empirical evaluation of over ten distinct representation-based methods, spanning sequential, graph-based, and text-based architectures. Our central finding is a persistent trade-off between discrimination (AUROC) and calibration (ECE): text-based encoders achieve the best AUROC (0.672), while structurally-aware models yield the best ECE (0.148), with no single method dominating both. Furthermore, we find that increased architectural complexity does not reliably outperform simpler sequential baselines, suggesting a performance ceiling for methods relying solely on chunk-level hidden states. This work provides the most comprehensive benchmark for this task to date, establishing rigorous baselines and demonstrating the limitations of current representation-based paradigms.

With the advent of large language models (LLMs), it has become common practice for users to draft text and utilize LLMs to enhance its quality through paraphrasing. However, this process can sometimes result in the loss or distortion of the original intended meaning. Due to the human-like quality of LLM-generated text, traditional detection methods often fail, particularly when text is paraphrased to closely mimic original content. In response to these challenges, we propose a novel approach named SearchLLM, designed to identify LLM-paraphrased text by leveraging search engine capabilities to locate potential original text sources. By analyzing similarities between the input and regenerated versions of candidate sources, SearchLLM effectively distinguishes LLM-paraphrased content. SearchLLM is designed as a proxy layer, allowing seamless integration with existing detectors to enhance their performance. Experimental results across various LLMs demonstrate that SearchLLM consistently enhances the accuracy of recent detectors in detecting LLM-paraphrased text that closely mimics original content. Furthermore, SearchLLM also helps the detectors prevent paraphrasing attacks.

Large language models (LLMs) have achieved remarkable success across a wide range of tasks, yet fine-tuning them efficiently under black-box or memory-constrained settings remains challenging. Parameter-efficient fine-tuning (PEFT) techniques such as LoRA alleviate memory usage by restricting updates to low-rank adapters, while zeroth-order (ZO) optimization further avoids back-propagation by estimating gradients from function evaluations. Recent work, such as LOZO, leverages random low-rank perturbations to reduce the variance of ZO estimates, but it overlooks the intrinsic geometric structure of LoRA adapters and suffers from unstable convergence and limited integration with adaptive optimizers. To address these limitations, we propose RoZO, a Riemannian zeroth-order optimization framework that constrains updates to the tangent space of the LoRA manifold. By exploiting geometry-aware updates with parallel transport, adaptive preconditioning, and trust-region control, RoZO achieves more stable convergence, tighter variance bounds, and superior performance compared to existing ZO methods.

Entity representation learning on hypergraphs is hindered by degree bias, where nodes with sparse connections suffer from limited structural information for aggregation. Prevailing “attribute-as-feature“ approaches, which treat rich textual attributes (e.g., titles, abstracts, keywords) merely as node features, fail to address this structurally rooted problem as they do not create new aggregation pathways. To overcome this limitation, we propose a novel “attribute-as-structure“ approach specifically designed for heterogeneous hypergraphs. Our approach integrates attributes directly into the hypergraph topology as distinct node types, creating new structural pathways to enrich sparsely connected entities while preserving semantic distinctiveness within complex many-to-many hyperedge interactions. We introduce an entity-attribute aware learning framework featuring two key innovations: (1) a specialized heterogeneous hypergraph encoder with dual attention mechanisms—self-attention for entity-entity relationships and cross-type attention for entity-attribute relevance, and (2) Attribute-Attentive Contrastive Learning (AACL), a novel objective that dynamically weighs attribute importance while explicitly aligning entity representations with their structural attributes. Experiments on multiple hypergraph datasets demonstrate consistent improvements in node classification performance, with particularly significant gains for structurally sparse nodes, demonstrating the effectiveness of our approach for degree bias mitigation.

Despite their potential as human proxies, LLMs often fail to generate heterogeneous data with human-like diversity, thereby diminishing their value in advancing social science research. To address this gap, we propose a novel method to incorporate psychological insights into LLM simulation through the Personality Structured Interview (PSI). PSI leverages psychometric scale-development procedures to capture personality-related linguistic information from a formal psychological perspective. To systematically evaluate simulation fidelity, we developed a measurement theory grounded evaluation procedure that considers the latent construct nature of personality and evaluates its reliability, structural validity, and external validity. Results from three experiments demonstrate that PSI effectively improves human-like heterogeneity in LLM-simulated personality data and predicts personality-related behavioral outcomes. We further offer a theoretical framework for designing theory-informed structured interviews to enhance the reliability and effectiveness of LLMs in simulating human-like data for broader psychometric research.

Large Language Models (LLMs) have achieved substantial progress in alignment, ensuring safer and more reliable outputs. However, jailbreak attacks can still bypass these safeguards and provoke harmful responses from well-aligned models. While some studies have achieved defenses against jailbreak attacks by modifying output distributions or detecting harmful content, the exact rationale still remains elusive. In this work, we present a novel neuron-level interpretability method that focuses on the role of safety-related knowledge neurons. Unlike existing approaches, our method projects the model’s internal representation into a more consistent and interpretable vocabulary space. We then show that adjusting the activation of safety-related neurons can effectively control the model’s behavior with a mean ASR higher than 97%. Building on this insight, we propose SafeTuning, a fine-tuning strategy that reinforces safety-critical neurons to improve model robustness against jailbreaks. SafeTuning consistently reduces attack success rates across multiple LLMs and outperforms all four baseline defenses. These findings offer a new perspective on understanding and defending against jailbreak attacks.

Large Language Models (LLMs) suffer from severe catastrophic forgetting when adapted sequentially to new tasks in a continual learning (CL) setting. Existing approaches are fundamentally limited: replay-based methods are impractical and could potentially violate privacy, while strict orthogonality-based methods collapse under scale: each new task is projected onto an orthogonal complement, progressively reducing the residual degrees of freedom and eliminating forward transfer by forbidding overlap in shared representations. In this work, we introduce ELLA, a training framework built on the principle of selective subspace de-correlation. Rather than forbidding all overlap, ELLA explicitly characterizes the structure of past updates and penalizes alignments along their high-energy, task-specific directions, while preserving freedom in the low-energy residual subspaces to enable transfer. Formally, this is realized via a lightweight regularizer on a single aggregated update matrix. This mechanism is proven to be an anisotropic shrinkage operator that bounds interference, yielding a penalty that is both memory- and compute-constant regardless of task sequence length. ELLA requires no data replay, no architectural expansion, and negligible storage. Empirically, it achieves state-of-the-art CL performance on three popular benchmarks spanning both classification and generative tasks, with relative accuracy gains of up to 9.6% and a 35× smaller memory footprint. Furthermore, ELLA scales robustly across architectures and actively enhances the model’s zero-shot generalization performance on unseen tasks, establishing a principled and scalable solution for constructive lifelong LLM adaptation.

We present LingGen, a controlled text generation model that allows fine-grained control over a large number of real-valued linguistic attributes. It encodes target attribute values with a dedicated linguistic attribute encoder and conditions the language model by injecting the resulting representation into the language model using the beginning-of-sequence (BOS) embeddings. To improve robustness when controlling different attribute subsets, we introduce P-MASKING, which samples per-example attribute masking rates from a truncated Pareto distribution during training. Across 1-40 control attributes, LingGen achieves the lowest average control error among evaluated methods, while remaining efficient at inference and receiving the highest fluency scores in human evaluation. Ablations show that Pareto-sampled masking and BOS-based injection are effective choices compared to alternative masking and integration variants.

Temporal Knowledge Graphs (TKGs) represent dynamic facts as timestamped relations between entities. While Large Language Models (LLMs) show promise for TKG completion, current approaches typically apply generic pipelines (neighborhood sampling, supervised fine-tuning, uncalibrated inference) without task-specific adaptation to temporal relational reasoning. Through systematic analysis under unified evaluation, we reveal three key failure modes: (1) retrieval strategies miss multi-hop dependencies when target entities are not directly observed in history, (2) standard fine-tuning reinforces memorization over relational generalization, and (3) uncalibrated generation produces contextually implausible entities. We present RECIPE-TKG, a parameter-efficient framework that addresses each limitation through principled, task-specific design: rule-based multi-hop sampling for structural grounding, contrastive fine-tuning to shape relational compatibility, and test-time semantic filtering for contextual alignment. Experiments on four benchmarks show that RECIPE-TKG outperforms prior LLM-based methods across input regimes, achieving up to 22.4% relative improvement in Hits@10, with particularly strong gains when historical evidence is sparse or indirect.

Transformers have become the foundational architecture for a broad spectrum of sequence modeling applications, underpinning state-of-the-art systems in natural language processing, vision, and beyond. However, their theoretical limitations in discrete reasoning tasks, such as arithmetic, logical inference, and algorithmic composition, remain a critical open problem. In this survey, we synthesize recent advances from three theoretical perspectives: circuit complexity, approximation theory, and communication complexity, to clarify the structural and computational barriers that transformers face when performing symbolic computations. By connecting these established theoretical frameworks, we provide an accessible and unified account of why current transformer architectures struggle to implement exact discrete algorithms, even as they excel at pattern matching and interpolation. We review key definitions, seminal results, and illustrative examples, highlighting challenges such as depth constraints, difficulty approximating discontinuities, and bottlenecks in inter-token communication. Finally, we discuss implications for model design and suggest promising directions for overcoming these foundational limitations.

Search agents are language models (LMs) that reason and search knowledge bases (or the web) to answer questions; recent methods supervise only the final answer accuracy using reinforcement learning with verifiable rewards (RLVR). Most RLVR search agents tackle general-domain QA, which limits their relevance to technical AI systems in science, engineering, and medicine. In this work we propose training agents to search and reason over scientific papers – this tests technical question-answering, it is directly relevant to real scientists, and the capabilities will be crucial to future AI Scientist systems. Concretely, we release a search corpus of 16 million biomedical paper abstracts and construct a challenging factoid QA dataset called PaperSearchQA with 60k samples answerable from the corpus, along with benchmarks. We train search agents in this environment to outperform non-RL retrieval baselines; we also perform further quantitative analysis and observe interesting agent behaviors like planning, reasoning, and self-verification. Our corpus, datasets, and benchmarks are usable with the popular Search-R1 codebase for RLVR training; they are available on Hugging Face. Finally, our data creation methods are scalable and easily extendable to other scientific domains.

Large Language Models (LLMs) are increasingly used to simulate public opinion and other social phenomena. Most current studies constrain these simulations to multiple-choice or short-answer formats for ease of scoring and comparison, but such closed designs overlook the inherently generative nature of LLMs. In this position paper, we argue that open-endedness, using free-form text that captures topics, viewpoints, and reasoning processes "in" LLMs, is essential for realistic social simulation. Drawing on decades of survey methodology research and recent advances in NLP, we argue why this open-endedness is valuable in LLM social simulations, showing how it can improve measurement and design, support exploration of unanticipated views, and reduce researcher-imposed directive bias. It also captures expressiveness and individuality, aids in pretesting, and ultimately enhances methodological utility. We call for novel practices and evaluation frameworks that leverage rather than constrain the open-ended generative diversity of LLMs, creating synergies between NLP and social science.

Retrieval-augmented generation (RAG) based on large language models often falters on narrative documents with inherent temporal structures. Standard unstructured RAG methods rely solely on embedding-similarity matching and lack any general mechanism to encode or exploit chronological information, while knowledge graph RAG (KG-RAG) frameworks collapse every mention of an entity into a single node, erasing the evolving context that drives many queries. To formalize this challenge and draw the community’s attention, we construct ChronoQA, a robust and discriminative QA benchmark that measures temporal, causal, and character consistency understanding in narrative documents (e.g., novels) under the RAG setting. We then introduce Entity-Event RAG (E ²RAG), a dual-graph framework that keeps separate entity and event subgraphs linked by a bipartite mapping, thereby preserving the temporal and causal facets needed for fine-grained reasoning. Across ChronoQA, our approach outperforms state-of-the-art unstructured and KG-based RAG baselines, with notable gains on causal and character consistency queries. E ²RAG therefore offers a practical path to more context-aware retrieval for tasks that require precise answers grounded in chronological information.

The advent of Large Language Models (LLMs) has significantly advanced web-based Question Answering (QA) systems over semi-structured content, raising questions about the continued utility of knowledge extraction for question answering. This paper investigates the value of triple extraction in this new paradigm by extending an existing benchmark with knowledge extraction annotations and evaluating commercial and open-source LLMs of varying sizes. Our results show that web-scale knowledge extraction remains a challenging task for LLMs. Despite achieving high QA accuracy, LLMs can still benefit from knowledge extraction, through augmentation with extracted triples and multi-task learning. These findings provide insights into the evolving role of knowledge triple extraction in web-based QA and highlight strategies for maximizing LLM effectiveness across different model sizes and resource settings.

Images often communicate more than they literally depict: a set of tools can suggest an occupation and a cultural artifact can suggest a tradition. This kind of indirect visual reference, known as visual metonymy, invites viewers to recover a target concept via associated cues rather than explicit depiction. In this work, we present the first computational investigation of visual metonymy. We introduce a novel pipeline grounded in semiotic theory that leverages large language models and text-to-image models to generate metonymic visual representations. Using this framework, we construct ViMET, the first visual metonymy dataset comprising 2,000 multiple-choice questions to evaluate the cognitive reasoning abilities in multimodal language models. Experimental results on our dataset reveal a significant gap between human performance (86.9%) and state-of-the-art vision-language models (65.9%), highlighting limitations in machines’ ability to interpret indirect visual references. Our dataset is publicly available at: https://github.com/cincynlp/ViMET.

Judeo-Arabic refers to Arabic variants historically spoken by Jewish communities across the Arab world, primarily during the Middle Ages. Unlike standard Arabic, it is written in Hebrew script by Jewish writers and for Jewish audiences. Transliterating Judeo-Arabic into Arabic script is challenging due to ambiguous letter mappings, inconsistent orthographic conventions, and frequent code-switching into Hebrew. In this paper, we introduce a two-step approach to automatically transliterate Judeo-Arabic into Arabic script: simple character-level mapping followed by post-correction to address grammatical and orthographic errors. We also present the first benchmark evaluation of LLMs on this task. Finally, we show that transliteration enables Arabic NLP tools to perform morphosyntactic tagging and machine translation, which would have not been feasible on the original texts. We make our code and data publicly available.

Multimodal large language models (MLLMs) are currently at the center of research attention, showing rapid progress in scale and capabilities, yet their intelligence, limitations, and risks remain insufficiently understood. To address these issues, particularly in the context of the Russian language, where no multimodal benchmarks currently exist, we introduce MERA Multi, an open multimodal evaluation framework for Russian-spoken architectures. The benchmark is instruction-based and encompasses default text, image, audio, and video modalities, comprising 18 newly constructed evaluation tasks for both general-purpose models and modality-specific architectures (image-to-text, video-to-text, and audio-to-text). Our contributions include: (i) a universal taxonomy of multimodal abilities; (ii) 18 datasets created entirely from scratch with attention to Russian cultural and linguistic specificity, unified prompts, and metrics; (iii) baseline results for both closed-source and open-source models; (iv) a methodology for preventing benchmark leakage, including watermarking for private sets. While our current focus is on Russian, the proposed benchmark provides a replicable methodology for constructing multimodal benchmarks in typologically diverse languages, particularly within the Slavic language family.

Tasks such as solving arithmetic equations, evaluating truth tables, and completing syllogisms are handled well by large language models (LLMs) in their standard form, but they often fail when the same problems are posed in logically equivalent yet obfuscated formats. To study this vulnerability, we introduce Logifus, a structure-preserving logical obfuscation framework, and, utilizing this, we present LogiQAte, a first-of-its-kind diagnostic benchmark with 1,108 questions across four reasoning tasks: (i) Obfus FOL (first-order logic entailment under equivalence-preserving rewrites), (ii) Obfus Blood Relation (family-graph entailment under indirect relational chains), (iii) Obfus Number Series (pattern induction under symbolic substitutions), and (iv) Obfus Direction Sense (navigation reasoning under altered directions and reference frames). Across all the tasks, evaluating six state-of-the-art models, we find that obfuscation severely degrades zero-shot performance, with performance dropping on average by 47% for GPT-4o, 27% for GPT-5, and 22% for reasoning model, o4-mini. Our findings reveal that current LLMs parse questions without deep understanding, highlighting the urgency of building models that genuinely comprehend and preserve meaning beyond surface form.

Despite advances in large language models (LLMs), Task-Oriented Dialogue (TOD) systems often fall short in delivering personalized, context-rich responses, especially in low-resource, code-mixed, and multimodal settings like Hinglish (Hindi-English). To bridge this gap, we introduce HiVisTask, the first Hinglish multimodal, multidomain, persona-based TOD dataset that captures user-agent interactions across text and visual modalities. We also propose G³ TOD, a generalizable framework that enhances personalization using three structured knowledge graphs: entity context, user persona, and commonsense reasoning, all extracted from conversation history. Extensive experiments with LLMs (e.g., LLaMA3.2, Phi3, GPT4, Mistral7b, Qwen3, Gemma3) show that G³ TOD consistently outperforms both standard and ablated baselines. We observe substantial gains across evaluation metrics (both quantitative: BLEU ↑ and qualitative: Human Eval ↑) over existing models. The observed improvements strongly underscore the value of structured and selective contextualization in generating personalized and engaging multimodal responses.

Chain-of-Thought (CoT) prompting has emerged as a powerful technique for enhancing language model’s reasoning capabilities. However, generating long and correct CoT trajectories is challenging. Recent studies have demonstrated that Looped Transformers, a standard Transformer with cross-block parameter-sharing architecture, possess remarkable length generalization capabilities, but their limited generality and adaptability prevent them from serving as an alternative to auto-regressive solutions. To better leverage the strengths of Looped Transformers, we propose **RELAY** (**RE**asoning through **L**oop **A**lignment iterativel**Y**). Specifically, we align the steps of Chain-of-Thought (CoT) reasoning with loop iterations and apply intermediate supervision during the training of Looped Transformers. This additional iteration-wise supervision not only preserves the Looped Transformer’s ability for length generalization but also enables it to predict CoT reasoning steps for unseen data. Therefore, we leverage this Looped Transformer to generate accurate reasoning chains for complex problems that exceed the training length, which will then be used to fine-tune an auto-regressive model. We conduct extensive experiments, and the results demonstrate the effectiveness of our approach, with significant improvements in the performance of the auto-regressive model.

Despite the rapid advancements of Large Language Models (LLMs), safety risks remain a critical challenge for low-resource languages. Existing safety datasets are predominantly English-centric, limiting progress in multilingual safety alignment. As a result, low-resource expert models—fine-tuned on their respective instruction datasets—tend to exhibit higher unsafety rates compared to their high-resource counterparts. In this work, we propose a safety aware layer swapping method that transfers safety alignment from an English safety expert to low-resource language experts without additional training. To further enhance transfer ability, our method adaptively selects or blends modules based on their degree of specialization. Our approach preserves performance on general language understanding tasks while enhancing safety in the target languages. Experimental results show that the proposed method achieves comparable performance to the language expert on general benchmarks such as MMMLU, BELEBELE, and MGSM, while producing more aligned and less harmful responses on the MultiJail safety benchmark

The principle of compositionality, which concerns the construction of meaning from constituent parts, is a longstanding topic in various disciplines, most commonly associated with formal semantics. In NLP, recent studies have focused on the compositional properties of text embedding models, particularly regarding their sensitivity to idiomatic expression, as idioms have traditionally been seen as non-compositional. In this paper, we argue that it is unclear how previous work relates to formal definitions of the principle. To address this limitation, we take a theoretically motivated approach based on definitions in formal semantics. We present 𝜀-compositionality, a continuous relaxation of compositionality derived from these definitions. We measure 𝜀-compositionality on a dataset containing both idiomatic and non-idiomatic sentences, providing a theoretically motivated assessment of sensitivity to idiomaticity. Our findings indicate that most text embedding models differentiate between idiomatic and non-idiomatic phrases, although to varying degrees.

Large language model (LLM) systems increasingly power everyday AI applications such as chatbots, computer-use assistants, and autonomous robots, where performance often depends on manually well-crafted prompts. LLM-based prompt optimizers reduce that effort by iteratively refining prompts from scored feedback, yet the security of this optimization stage remains underexamined. We present the first systematic analysis of poisoning risks in LLM-based prompt optimization. Using HarmBench, we find systems are substantially more vulnerable to manipulated feedback than to query poisoning alone: feedback-based attacks raise attack success rate (ASR) by up to ΔASR = 0.48. We introduce a simple fake reward attack that requires no access to the reward model and significantly increases vulnerability. We also propose a lightweight highlighting defense that reduces the fake reward ΔASR from 0.23 to 0.07 without degrading utility. These results establish prompt optimization pipelines as a first-class attack surface and motivate stronger safeguards for feedback channels and optimization frameworks.

Despite recent advances, long-sequence video generation frameworks still suffer from significant limitations: poor assistive capability, suboptimal visual quality, and limited expressiveness. To mitigate these limitations, we propose MAViS, an end-to-end multi-agent collaborative framework for long-sequence video storytelling. MAViS orchestrates specialized agents across multiple stages, including script writing, shot designing, character modeling, keyframe generation, video animation, and audio generation. In each stage, agents operate under the 3E Principle—Explore, Examine, and Enhance—to ensure the completeness of intermediate outputs. Considering the capability limitations of current generative models, we propose the Script Writing Guidelines to optimize compatibility between scripts and generative tools. Experimental results demonstrate that MAViS achieves state-of-the-art performance in assistive capability, visual quality, and video expressiveness. Its modular framework further enables scalability with diverse generative models and tools. With just a brief prompt, MAViS enables users to rapidly explore diverse visual storytelling and creative directions for sequential video generation by efficiently producing high-quality, complete long-sequence videos. To the best of our knowledge, MAViS is the only framework that provides multimodal design output – videos with narratives and background music.

We present AraBabyTalk-EGY, an enriched release of the Egyptian Arabic CHILDES corpus, that opens the child-adult interactions genre to modern Arabic NLP research. Starting from the original CHILDES recordings and IPA transcriptions of caregiver-child sessions, we (i) map each IPA token to fully diacritized Arabic script, and (ii) add core part-of-speech tags and lemmas aligned with existing dialectal Arabic morphological resources. These layers yield ~26K annotated tokens suitable for both text- and speech-based NLP tasks. We provide a benchmark on morphological disambiguation and Arabic ASR. We outline lexical and morphosyntactic differences between AraBabyTalk-EGY and general Egyptian Arabic resources, highlighting the value of genre-specific training data for language acquisition studies and Arabic speech technology.

The Korean legal system mandates public access to court judgments to ensure judicial transparency. However, this requirement conflicts with privacy protection obligations due to the prevalence of Personally Identifiable Information (PII) in legal documents. To address this challenge, we introduce **K-LegalDeID**, a large-scale benchmark dataset and an efficient KLUEBERT-CRF model for de-identification for Korean court judgments. Our primary contribution is a new large-scale benchmark dataset spanning 39 legal domains, with its quality is validated by a high inter-annotator agreement (IAA) with Fleiss’ Kappa of 0.7352. Our results demonstrate that a lightweight KLUEBERT-CRF model, when trained on our dataset, achieves state-of-the-art performance with an entity-level micro F1 score of 0.9923. Our end-to-end framework offers a practical and computationally efficient solution for real-world legal systems.

Mixture-of-Experts (MoE) based large language models (LLMs) have gained popularity due to their multi-task capability, where each input token activates only a subset of "expert" subnetworks. However, whether each expert can truly specialize to a certain task remains poorly understood, while activation analysis shows frequent cross-layer co-activation of experts for the same input, resembling a collaborative behavior. In this paper, we use a dictionary learning approach to show that experts in MoE LLMs form hierarchical and semantically coherent collaborative groups that correspond to specific linguistic and cognitive functions (e.g., mathematical reasoning, syntactic processing), mirroring specialized functional region observed in neuroscience. Furthermore, leveraging these discovered expert groups enables significant model compression with minimal performance degradation, outperforming existing methods by 2.5% while enabling up to 50% expert reduction. These findings provide the first systematic analysis of expert collaboration mechanisms in MoE LLMs, revealing that specialization emerges from joint activation of experts across all layers. We further developed an interactive visualization platform that enables researchers to explore expert collaboration patterns and their semantic associations.

Health wellness agents typically rely on large language models (LLMs) for response generation, where contextual information from a user’s conversation history can be used for response grounding and personalization. High-quality conversation summaries are one such method which can reduce the number of input tokens during response generation, decreasing overhead and inference latency. However, directly purposing LLMs for this task is infeasible due to the scale of the task, the compute overhead, and health data compliance regulations. Furthermore, ground truth for real-world datasets is scarce due to privacy concerns and the high cost of health expert annotators. These factors necessitate the development of small, potentially on-device, language models capable of health dialogue summarization, particularly in the absence of ground truth labels. In this paper, we first present a comprehensive empirical study that benchmarks a variety of state-of-the-art smaller language models to better understand their baseline capabilities. Second, we present an unsupervised method that uses the summaries from multiple models, refined with iterative feedback, to generate high-quality summaries of health dialogues. Experiments illustrate that our method is able to outperform baseline on both open-source and proprietary benchmarks. Notably, our method can be run viably on local compute without a GPU, using just a single Macbook with 16 GB of memory.

Reliable uncertainty quantification (UQ) is essential when employing large language models (LLMs) in high-risk domains such as clinical question answering (QA). In this work, we evaluate uncertainty estimation methods for clinical QA focusing, for the first time, on eleven clinical specialties and six question types, and across ten open-source LLMs (general-purpose, biomedical, and reasoning models), alongside representative proprietary models. We analyze score-based UQ methods, present a case study introducing a novel lightweight method based on behavioral features derived from reasoning-oriented models, and examine conformal prediction as a complementary set-based approach. Our findings reveal that uncertainty reliability is not a monolithic property, but one that depends on clinical specialty and question type due to shifts in calibration and discrimination. Our results highlight the need to select or ensemble models based on their distinct, complementary strengths and clinical use.

The way our eyes move while reading can tell us about the cognitive effort required to process the text. In the present study, we use this fact to generate texts with controllable reading ease. Our method employs a model that predicts human gaze patterns to steer language model outputs towards eliciting certain reading behaviors. We evaluate the approach in an eye-tracking experiment with native and non-native speakers of English. The results demonstrate that the method is effective at making the generated texts easier or harder to read, measured both in terms of reading times and perceived difficulty of the texts. A statistical analysis reveals that the changes in reading behavior are mostly due to features that affect lexical processing. Possible applications of our approach include text simplification for information accessibility and generation of personalized educational material for language learning.

We investigate the automated evaluation of English language learner answers to writing tasks featuring picture stories.This is usually limited to language proficiency only, neglecting the context of the picture. Instead, our analysis focuses on task adherence, which for example allows detection of off-topic answers.Since there is a lack of suitable training and evaluation data, our first step is to build the PictureStories dataset.To this end, we develop a marking rubric that covers task adherence with respect to both form and content. Six annotators mark 713 learner answers written in response to one of five picture stories.Having assembled the dataset, we then explore to what extent task adherence can be predicted automatically. Our experiments assume a scenario where no or just a few labelled answers are available for the picture story which is being marked.For form-focused criteria, we find that it is beneficial to finetune models across tasks.With content-focused criteria, few-shot prompting Qwen emerges as the best-performing method. We examine the trade-off between including the story image vs. example answers in the prompt and find that examples suffice in many cases. While for some LLMs, few-shot prompting results may look promising on the surface, we demonstrate that a much simpler method can do just as well when shown the same examples.

Despite major advances in multilingual modeling, large quality disparities persist across languages. Besides the obvious impact of uneven training resources, typological properties have also been proposed to determine the intrinsic difficulty of modeling a language. The existing evidence, however, is mostly based on small monolingual language models or bilingual translation models trained from scratch. We expand on this line of work by analyzing two large pre-trained multilingual translation models, NLLB-200 and Tower+, which are state-of-the-art representatives of encoder-decoder and decoder-only machine translation, respectively. Based on a broad set of languages, we find that target language typology drives translation quality of both models, even after controlling for more trivial factors, such as data resourcedness and writing script. Additionally, languages with certain typological properties benefit more from a wider search of the output space, suggesting that such languages could profit from alternative decoding strategies beyond the standard left-to-right beam search. To facilitate further research in this area, we release a set of fine-grained typological properties for 212 languages of the FLORES+ MT evaluation benchmark.

There are many methods and systems to tackle the ontology alignment problem, yet a major challenge persists in producing high-quality mappings among a set of input ontologies. Adopting a human-in-the-loop approach during the alignment process has become essential in applications requiring very accurate mappings. However, user involvement is expensive when dealing with large ontologies. In this paper, we analyse the feasibility of using Large Language Models (LLM) to aid the ontology alignment problem. LLMs are used only in the validation of a subset of correspondences for which there is high uncertainty. We have conducted an extensive analysis over several tasks of the Ontology Alignment Evaluation Initiative (OAEI), reporting in this paper the performance of several state-of-the-art LLMs using different prompt templates. Using LLMs as resulted in strong performance in the OAEI 2025, achieving the top-2 overall rank in the bio-ml track.

Medical reasoning in large language models seeks to replicate clinicians’ cognitive processes in interpreting patient data and making diagnostic decisions. However, widely used benchmarks—such as MedQA, MedMCQA, and PubMedQA—mix questions that require multi-step reasoning with those answerable through factual recall, complicating evaluation. We introduce an expert-validated evaluation framework that disentangles knowledge recall from reasoning by training a PubMedBERT-based classifier and applying it to 11 widely used biomedical QA benchmarks. This framework reveals that only 32.8% of questions require multi-step reasoning, indicating that current evaluations largely measure factual recall. Stratified evaluation of biomedical models (HuatuoGPT-o1, MedReason, m1) and general-domain models (DeepSeek-R1, o4-mini, Qwen3) consistently shows lower performance on reasoning-heavy than knowledge-heavy questions (e.g., HuatuoGPT-o1: 56.9% on knowledge vs.44.8% on reasoning). Beyond aggregate accuracy, we assess robustness through adversarial evaluations in which models are prefixed with uncertainty-inducing, incorrect statements; biomedical reasoning models degrade sharply in this setting (e.g., MedReason: 50.4% to 24.4%), with declines especially pronounced on reasoning-heavy questions. Finally, we show that fine-tuning on high-quality, reasoning-heavy examples augmented with adversarial traces, followed by reinforcement learning with GRPO, improves both robustness and accuracy across knowledge and reasoning subsets within our evaluation framework.

Explicit representations of predicate-argument relations form the basis of interpretable semantic analysis, supporting reasoning, generation, and evaluation. However, attaining such semantic structures requires costly annotation efforts and has remained largely confined to English. We leverage the Question-Answer driven Semantic Role Labeling (QA-SRL) framework — a natural-language formulation of predicate-argument relations — as the foundation for extending semantic annotation to new languages. To this end, we introduce a cross-linguistic projection approach that reuses an English QA-SRL parser within a constrained translation and word-alignment pipeline to automatically generate question-answer annotations aligned with target-language predicates. Applied to Hebrew, Russian, and French — spanning diverse language families — the method yields high-quality training data and fine-tuned, language-specific parsers that outperform strong multilingual LLM baselines (GPT-4o, LLaMA-Maverick). By leveraging QA-SRL as a transferable natural-language interface for semantics, our approach enables efficient and broadly accessible predicate-argument parsing across languages.

Intrinsic evaluation metrics for conditional language models, such as perplexity or bits-per-character, are widely used in both mono- and multilingual settings. These metrics are rather straightforward to use and compare in monolingual setups, but rest on a number of assumptions in multilingual setups. One such assumption is that comparing the perplexity of CLMs on parallel sentences is indicative of their quality since the information content (here understood as the semantic meaning) is the same. However, the metrics are inherently measuring information content in the information-theoretic sense. We make this and other such assumptions explicit and discuss their implications. We perform experiments with six metrics on two multi-parallel corpora both with mono- and multilingual models. Ultimately, we find that current metrics are not universally comparable. We look at the form-meaning debate to provide some explanation for this.

Knowledge Graph-based Retrieval-Augmented Generation (KG-RAG) is an increasingly explored approach for combining the reasoning capabilities of large language models with the structured evidence of knowledge graphs. However, current evaluation practices fall short: existing benchmarks often include questions that can be directly answered using existing triples in KG, making it unclear whether models perform reasoning or simply retrieve answers directly. Moreover, inconsistent evaluation metrics and lenient answer matching criteria further obscure meaningful comparisons. In this work, we introduce a general method for constructing benchmarks and present BRINK (Benchmark for Reasoning under Incomplete Knowledge) to systematically assess KG-RAG methods under knowledge incompleteness. Our empirical results show that current KG-RAG methods have limited reasoning ability under missing knowledge, often rely on internal memorization, and exhibit varying degrees of generalization depending on their design.

Chat assistants increasingly integrate web search functionality, enabling them to retrieve and cite external sources. While this promises more reliable answers, it also raises the risk of amplifying misinformation from low-credibility sources. In this paper, we introduce a novel methodology for evaluating assistants’ web search behavior, focusing on source credibility and the groundedness of responses with respect to cited sources. Using 100 claims across five misinformation-prone topics, we assess GPT-4o, GPT-5, Perplexity, and Qwen Chat. Our findings reveal differences between the assistants, with Perplexity achieving the highest source credibility, whereas GPT-4o exhibits elevated citation of non-credible sources on sensitive topics. This work provides the first systematic comparison of commonly used chat assistants for fact-checking behavior, offering a foundation for evaluating AI systems in high-stakes information environments.

Large Language Models (LLMs) need to be in accordance with human values—being helpful, harmless, and honest (HHH)—is important for safe deployment. Existing works use Supervised Fine-Tuning (SFT) and Mixture-of-Experts (MoE) to align LLMs. However, these works face challenges in multi-objective settings, such as SFT leading to interference between conflicting objectives, while MoEs suffer from miscalibrated routing. We term this failure mode Axis Collapse, marked by(1) disjoint feature spaces causing catastrophic forgetting, and (2) unreliable inference from misrouted experts. To resolve this, we propose AlignX, a two-stage framework. Stage 1 uses prompt-injected fine-tuning to extract axis-specific task features, mitigating catastrophic forgetting. Stage 2 deploys a MoCaE module that calibrates expert routing using fractal and natural geometry, improving inference reliability. AlignX achieves significant gains on Alpaca (Helpfulness), BeaverTails (Harmlessness), and TruthfulQA (Honesty), with +171.5% win rate, +110.1% in truthfulness-informativeness, and 4.3% fewer safety violations. It also reduces latency and memory usage by over 35% compared to prior MoEs. Results across four LLMs validate its generalizability. Code and data are available at: https://github.com/gskgautam/AlignX

Extending large language models to low-resource languages is essential for global accessibility, but training separate models per language is prohibitively expensive. Mixture-of-Experts (MoE) architectures address this by adding sparse language-specific parameters, but determining how many experts each layer needs remains an open question. Current approaches allocate experts based on layer-level similarity, yet language processing exhibits fine-grained specialization at individual neurons. We propose NeuronMoE, a method that analyzes language-specific neurons across all transformer components to guide expert allocation per layer based on empirically measured cross-lingual neuron diversity. Applied to Llama-3.2-3B for low-resource languages (Greek, Turkish, and Hungarian), this approach achieves approximately 40% average parameter reduction while matching the performance of the LayerMoE baseline. We find that low-resource language experts independently develop neuron specialization patterns mirroring the high-resource language, which are concentrated in early and late layers. This reveals potential universal architectural principles in how multilingual models organize linguistic knowledge. Our approach generalizes across architectures, as validated on Qwen, and shows that allocation strategy matters more than total expert count.

Few forces rival fear in their ability to mobilize societies, distort communication, and reshape collective behavior. In computational linguistics, fear is primarily studied as an emotion, but not as a distinct form of speech. Fear speech content is widespread and growing, and often outperforms hate-speech content in reach and engagement because it appears "civiler" and evades moderation. Yet the computational study of fear speech remains fragmented and under-resourced. This can be understood by recognizing that fear speech is a phenomenon shaped by contributions from multiple disciplines. In this paper, we bridge cross-disciplinary perspectives by comparing theories of fear from Psychology, Political science, Communication science, and Linguistics. Building on this, we review existing definitions. We follow up with a survey of datasets from related research areas and propose a taxonomy that consolidates different dimensions of fear for studying fear speech. By reviewing current datasets and defining core concepts, our work offers both theoretical and practical guidance for creating datasets and advancing fear speech research.

Subword tokenization methods, such as Byte-Pair Encoding (BPE), significantly impact the performance and efficiency of large language models (LLMs). The standard approach involves training a general-purpose tokenizer that uniformly processes all textual data during both training and inference. However, the use of a generic set of tokens can incur inefficiencies when applying the model to specific domains or languages. To address this limitation, we propose a post-training adaptation strategy that selectively replaces low-utility tokens with more relevant ones based on their frequency in an adaptation corpus. Our algorithm identifies the token inventory that most effectively encodes the adaptation corpus for a given target vocabulary size. Extensive experiments on generation and classification tasks across multiple languages demonstrate that our adapted tokenizers compress test corpora more effectively than baselines using the same vocabulary size. This method serves as a lightweight adaptation mechanism, akin to a vocabulary fine-tuning process, enabling optimized tokenization for specific domains or tasks. Our code and data are available at https://github.com/vijini/Adapt-BPE.git.

Supervised learning relies on data annotation which usually is time-consuming and therefore expensive. A longstanding strategy to reduce annotation costs is active learning, an iterative process, in which a human annotates only data instances deemed informative by a model. Research in active learning has made considerable progress, especially with the rise of large language models (LLMs). However, we still know little about how these remarkable advances have translated into real-world applications, or contributed to removing key barriers to active learning adoption. To fill in this gap, we conduct an online survey in the NLP community to collect previously intangible insights on current implementation practices, common obstacles in application, and future prospects in active learning. We also reassess the perceived relevance of data annotation and active learning as fundamental assumptions. Our findings show that data annotation is expected to remain important and active learning to stay highly relevant while benefiting from LLMs. Consistent with a community survey from over 15 years ago, three key challenges yet persist—setup complexity, uncertain cost reduction, and tooling—for which we propose alleviation strategies. We publish an anonymized version of the dataset.

Novelty assessment is a central yet understudied aspect of peer review, particularly in high-volume fields like NLP where reviewer capacity is increasingly strained. We present a structured approach for automated novelty evaluation that models expert reviewer behavior through three stages: (i) content extraction from submissions, (ii) retrieval and synthesis of related work, and (iii) structured comparison for evidence-based assessment. Our method is informed by analysis of human-written novelty reviews and captures key patterns such as independent claim verification and contextual reasoning. Evaluated on 182 ICLR 2025 submissions with human-annotated reviewer novelty assessments, the approach achieves 86.5% alignment with human reasoning and 75.3% agreement on novelty conclusions, substantially outperforming existing LLM-based baselines. It produces detailed, literature-aware analysis and improves consistency over ad hoc reviewer judgments. These results highlight the potential for structured LLM-assisted approaches to support more rigorous and transparent peer review without displacing human expertise. The data and the code are available at https://ukplab.github.io/eacl2026-assessing-paper-novelty/

Recent advances in multimodal and speech-native large language models (LLMs) have delivered impressive speech recognition, translation, understanding, and question-answering capabilities for high-resource languages. However, African languages and non-native French or English accents remain dramatically underrepresented in benchmarks limiting the understanding and applicability of leading LLMs for millions of francophone and anglophone users in low-resource settings. We presents AfriVox, an open-source benchmark (including novel domain-specific and unscripted datasets) across 20 African languages, African-accented French, Arabic, and 100+ African English accents, contrasting leading multimodal speech LLMs with traditional unimodal automatic speech transcription (ASR) and translation (AST) models. Our analysis reveals significant language coverage variation, surprising LLM translation performance gains (e.g. Gemini), robustness concerns with unscripted speech, and substantial performance disparities for "supported" African languages. We profile the strengths, limitations, and language support of each model, and conduct the first targeted fine-tuning of a modern speech LLM (Qwen2.5-Omni) for three Nigerian languages, exceeding SOTA, and achieving up to 54% relative WER reduction and significant BLEU gains, offering practical guidance for implementers seeking to serve local language users.

Historical language models play a crucial role in the study of languages, and can benefit tasks such as named-entity recognition (NER), part-of-speech (PoS) tagging, and post-OCR correction, among others. Despite their relevance, most efforts have been concentrated on English. To the best of our knowledge, no such model exists for historical Portuguese. In this work, we introduce PortOldBERT, the first historical Portuguese encoder language model. We demonstrate its usefulness by comparing PortOldBERT’s performance with Albertina, the encoder on which it is based, across multiple tasks—pseudo-perplexity, NER, PoS tagging, word error rate (WER) prediction, and OCR error detection—and for different historical periods. PortOldBERT consistently outperforms Albertina in historical data, demonstrating its ability to effectively integrate historical linguistic contexts while retaining the ability to process contemporary text.

Medical multiple-choice question answering (MCQA) benchmarks show that models achieve near-human accuracy, with some benchmarks approaching saturation–leading to claims of clinical readiness. Yet a single accuracy score is a poor proxy for competence: models that change answers under minor input perturbations cannot be considered reliable. We argue that reliability underpins accuracy–only consistent predictions make correctness meaningful. We release ReMedQA, a new benchmark that augments three standard medical MCQA datasets with open-ended answers and systematically perturbed options. Building on this, we introduce ReAcc and ReCon, two reliability metrics: ReAcc measures the proportion of questions answered correctly across all variations, while ReCon measures the proportion answered consistently regardless of correctness. Our evaluation shows that high MCQA accuracy masks low reliability: models remain sensitive to format and perturbation changes, and domain specialization offers no robustness gain. MCQA underestimates smaller models while inflating large ones that exploit structural cues–with some exceeding 50% accuracy even when the original questions are hidden. This shows that, despite near-saturated accuracy, we are not yet done with medical MCQA benchmarks.

Large Language Models (LLMs) often struggle with formal logical reasoning, frequently conflating content plausibility with logical validity. This well-known content effect undermines their capacity to act as reliable deductive reasoners, particularly in multilingual contexts where both linguistic variability and world knowledge may deepen biases. Prior work shows that prompting and tuning interventions can alleviate these issues only partially, leaving models vulnerable to semantic interference.While previous studies have explored activation steering and other test-time interventions, this work has focused predominantly on English.To make reasoning more consistent, robust, and transferable across languages, we investigate the use of activation steering—an inference-time intervention that modulates internal representations towards a cross-lingual reasoning space. Our experiments demonstrate that steering techniques constructed for English-based syllogisms generalise effectively to multilingual datasets, yielding higher formal reasoning accuracy (up to +36%) while minimally affecting language modelling performance. Moreover, steering supports partial transfer to out-of-distribution tasks, highlighting its potential as a scalable mechanism for cross-lingual transferable reasoning. These findings advance the prospect of developing LLMs that can serve as reliable soft reasoners across language landscapes.

Multimodal large language models (MLLMs) have shown impressive capabilities in vision-language tasks such as reasoning segmentation, where models generate segmentation masks based on textual queries. While prior work has primarily focused on perturbing image inputs, semantically equivalent textual paraphrases—crucial in real-world applications where users express the same intent in varied ways—remain underexplored. To address this gap, we introduce a novel adversarial paraphrasing task: generating grammatically correct paraphrases that preserve the original query meaning while degrading segmentation performance. To evaluate the quality of adversarial paraphrases, we develop a comprehensive automatic evaluation protocol validated with human studies. Furthermore, we introduce SPARTA—a black-box, sentence-level optimization method that operates in the low-dimensional semantic latent space of a text autoencoder, guided by reinforcement learning. SPARTA achieves significantly higher success rates, outperforming prior methods by up to 2x on both the ReasonSeg and LLMSeg-40k datasets. We use SPARTA and competitive baselines to assess the robustness of advanced reasoning segmentation models. We reveal that they remain vulnerable to adversarial paraphrasing—even under strict semantic and grammatical constraints. All code and data will be released publicly upon acceptance.

In this work, we propose an entity type-specific and knowledge-augmented token classification framework designed to improve encoder models’ performance on recipe texts. Our empirical analysis shows that this approach achieves state-of-the-art (SOTA) results on 5 out of 7 benchmark recipe datasets, significantly outperforming traditional token classification methods. We introduce a novel methodology leveraging curated domain-specific knowledge contexts to guide encoder models such as BERT and RoBERTa, which we refer to as RecipeBERT-KA and RecipeRoBERTa-KA. Additionally, we release a newly reprocessed entity type-specific and knowledge-enriched dataset that merges seven widely used food datasets, making it the largest annotated food-related dataset to date. Comparative analysis with SOTA large language models (GPT-4o, Mistral-7B, LLaMA 3-13B and LLaMA 3-70B) highlights the practical advantages of our smaller and specialised models. Finally, we analyse the impact of the different knowledge contexts, our models’ potential for transfer learning, the effect of combining the datasets and scenarios where traditional token classification may still perform competitively, offering nuanced insight into method selection.

In this work, we focus on discovery of legal factors for a specific case type under consideration (e.g., vehicle insurance disputes). We refer to these legal factors more explicitly as "Argumentation and Judgement Factors" (AJFs). AJFs encode specific legal knowledge that is important for legal argumentation and judicial decision making. We propose a multi-step approach for discovering a list of AJFs for a given case type using a set of relevant legal documents (e.g., past judgements, relevant acts) and Symbolic Knowledge Distillation (SKD) from a Large Language Model (LLM). We propose a novel geneRatE-CRitic-reviEW (RECREW) prompting strategy for effective SKD. We construct and evaluate the discovered list of AJFs on two different types of cases (auto-insurance and life insurance) and show their utility in a dispute resolution application.

Emotion classification plays a significant role in emotion prediction and harmful content detection. Recent advancements in NLP, particularly through large language models (LLMs), have greatly improved outcomes in this field. This study introduces ViGoEmotions - a Vietnamese emotion corpus comprising 20,664 social media comments in which each comment is classified into 27 fine-grained distinct emotions. To evaluate the quality of the dataset and its impact on emotion classification, eight pre-trained Transformer-based models were evaluated under three preprocessing strategies: preserving original emojis with rule-based normalization, converting emojis into textual descriptions, and applying ViSoLex, a model-based lexical normalization system. Results show that converting emojis into text often improves the performance of several BERT-based baselines, while preserving emojis yields the best results for ViSoBERT and CafeBERT. In contrast, removing emojis generally leads to lower performance. ViSoBERT achieved the highest Macro F1-score of 61.50% and Weighted F1-score of 63.26%. Strong performance was also observed from CafeBERT and PhoBERT. These findings highlight that while the proposed corpus can support diverse architectures effectively, preprocessing strategies and annotation quality remain key factors influencing downstream performance.

Current evaluations of sentence embedding models typically rely on static test beds such as the Massive Text Embedding Benchmark (MTEB). While invaluable, repeated tuning on a fixed suite can inflate reported performance and obscure real-world robustness. We introduce the Paraphrasing Text Embedding Benchmark (PTEB), a dynamic protocol that stochastically generates meaning-preserving paraphrases at evaluation time and aggregates results across multiple runs. Using a cost-efficient LLM-based method grounded in semantic textual similarity gold ratings, we show that LLMs generate token-diverse but semantically preserving, paraphrases. Across 7 MTEB tasks, we validate our hypothesis that the performance of sentence encoders is sensitive to changes in token space even when semantics remain fixed. We also observe that smaller models are not disproportionately affected relative to larger ones. Our results are statistically robust over multiple runs and we extended our experiments to 3 multilingual datasets covering 10 languages. More generally, we aim to propose a new evaluation paradigm in NLP that relies less on static, pre-defined benchmarks but shifts towards dynamic, stochastic evaluation leveraging eval-time compute. We make the code to run PTEB publicly available.

Current evaluation of German automatic text simplification (ATS) relies on general-purpose metrics such as SARI, BLEU, and BERTScore, which insufficiently capture simplification quality in terms of simplicity, meaning preservation, and fluency. While specialized metrics like LENS have been developed for English, corresponding efforts for German have lagged behind due to the absence of human-annotated corpora. To close this gap, we introduce DETECT, the first German-specific metric that holistically evaluates ATS quality across all three dimensions of simplicity, meaning preservation, and fluency, and is trained entirely on synthetic large language model (LLM) responses. Our approach adapts the LENS framework to German and extends it with (i) a pipeline for generating synthetic quality scores via LLMs, enabling dataset creation without human annotation, and (ii) an LLM-based refinement step for aligning grading criteria with simplification requirements. To the best of our knowledge, we also construct the largest German human evaluation dataset for text simplification to validate our metric directly. Experimental results show that DETECT achieves substantially higher correlations with human judgments than widely used ATS metrics, with particularly strong gains in meaning preservation and fluency. Beyond ATS, our findings highlight both the potential and the limitations of LLMs for automatic evaluation and provide transferable guidelines for general language accessibility tasks.

The increasing reliance on Large Language Models (LLMs) across various domains extends to education, where students progressively use generative AI as a tool for learning. While prior work has examined LLMs’ mathematical ability, their reliability in grading authentic student problem-solving processes and delivering effective feedback remains underexplored. This study introduces MathEDU, a dataset consisting of student problem-solving processes in mathematics and corresponding teacher-written feedback. We systematically evaluate the reliability of various models across three hierarchical tasks: answer correctness classification, error identification, and feedback generation. Experimental results show that fine-tuning strategies effectively improve performance in classifying correctness and locating erroneous steps. However, the generated feedback across models shows a considerable gap from teacher-written feedback. Critically, the generated feedback is often verbose and fails to provide targeted explanations for the student’s underlying misconceptions. This emphasizes the urgent need for trustworthy and pedagogy-aware AI feedback in education.

Test-time scaling (TTS) has enhanced the performance of Reasoning Models (RMs) on various tasks such as math and coding, yet its efficacy in machine translation (MT) remains underexplored. This paper investigates whether increased inference-time computation improves translation quality. We evaluate 12 RMs across a diverse suite of MT benchmarks spanning multiple domains, examining three scenarios: direct translation, forced-reasoning extrapolation, and post-editing. Our findings show that for general-purpose RMs, TTS provides limited and inconsistent benefits for direct translation, with performance quickly plateauing. However, the effectiveness of TTS is unlocked by domain-specific fine-tuning, which aligns a model’s reasoning process with task requirements, leading to consistent improvements up to an optimal, self-determined reasoning depth. We also find that forcing a model to reason beyond its natural stopping point consistently degrades translation quality. In contrast, TTS proves highly effective in a post-editing context, reliably turning self-correction into a beneficial process. These results indicate that the value of inference-time computation in MT lies not in enhancing single-pass translation with general models, but in targeted applications like multi-step, self-correction workflows and in conjunction with task-specialized models.

Most realistic task automation problems require large language models (LLMs) to call tools, which often return complex JSON responses. These responses must be further processed to derive the information necessary for task completion. The ability of LLMs to do so is under-studied. In this paper, we study the tool response processing task and LLMs’ abilities to process structured (JSON) responses. We created a dataset for this task, and evaluated 15 open and closed weight models using multiple prompting approaches. Our results show that JSON processing remains a difficult task even for frontier models across multiple prompting strategies. The optimal response processing strategy depends on both the nature and size of the tool outputs, as well as the complexity of the required reasoning. Variations in processing approaches can lead to performance differences ranging from 3% to 50%.

Idioms present a unique challenge for language models due to their non-compositional figurative interpretations, which often strongly diverge from the idiom’s literal interpretation. In this paper, we employ causal tracing to systematically analyze how pretrained causal transformers deal with this ambiguity. We localize three mechanisms: (i) Early sublayers and specific attention heads retrieve figurative interpretation, while suppressing literal interpretation. (ii) When disambiguating context precedes the idiom, the model leverages it from the earliest layer and later layers refine the interpretation if the context conflicts with the retrieved interpretation. (iii) Then, selective, competing pathways carry both interpretations: an intermediate pathway that prioritizes the figurative interpretation and a parallel direct route that favors literal interpretation, ensuring that both readings remain available. Our findings provide mechanistic evidence for idiom comprehension in autoregressive transformers.

LLMs, while outperforming humans in a wide range of tasks, can still fail in unanticipated ways. We focus on two pervasive failure modes: (i) hallucinations, where models produce incorrect information about the world, and (ii) the low-resource effect, where the models show impressive performance in high-resource languages like English but the performance degrades significantly in low-resource languages like Bengali. We study the intersection of these issues and ask: do hallucination detectors suffer from the low-resource effect? We conduct experiments on five tasks across three domains (factual recall, STEM, and Humanities). Experiments with four LLMs and three hallucination detectors reveal a curious finding: As expected, the task accuracies in low-resource languages experience large drops (compared to English). However, the drop in detectors’ accuracy is often several times smaller than the drop in task accuracy. Our findings suggest that even in low-resource languages, the internal mechanisms of LLMs might encode signals about their uncertainty. Further, the detectors are robust within language (even for non-English) and in multilingual setups, but not in cross-lingual settings without in-language supervision.

Rhetorical Role Labeling (RRL) identifies the functional role of each sentence in a document, a key task for discourse understanding in domains such as law and medicine. While hierarchical models capture local dependencies effectively, they are limited in modeling global, corpus-level features. To address this limitation, we propose two prototype-based methods that integrate local context with global representations. Prototype-Based Regularization (PBR) learns soft prototypes through a distance-based auxiliary loss to structure the latent space, while Prototype-Conditioned Modulation (PCM) constructs corpus-level prototypes and injects them during training and inference.Given the scarcity of RRL resources, we introduce SCOTUS-Law, the first dataset of U.S. Supreme Court opinions annotated with rhetorical roles at three levels of granularity: category, rhetorical function, and step. Experiments on legal, medical, and scientific benchmarks show consistent improvements over strong baselines, with ∼4 Macro-F1 gains on low-frequency roles. We further analyze the implications in the era of Large Language Models and complement our findings with expert evaluation.

Autoregressive (AR) models excel at generating temporally coherent audio by producing tokens sequentially, yet they often falter in faithfully following complex textual prompts—especially those describing complex sound events. We uncover a surprising capability in AR audio generators: their early prefix tokens implicitly encode global semantic attributes of the final output, such as event count and sound-object category, revealing a form of implicit planning. Building on this insight, we propose Plan-Critic, a lightweight auxiliary model trained with a Generalized Advantage Estimation (GAE)-inspired objective to predict final instruction-following quality from partial generations. At inference time, Plan-Critic enables guided exploration: it evaluates candidate prefixes early, prunes low-fidelity trajectories, and reallocates computation to high-potential planning seeds. Our Plan-Critic-guided sampling achieves up to a 10 points improvement in CLAP score over the AR baseline—establishing a new state of the art in AR text-to-audio generation—while maintaining computational parity with standard best-of-N decoding. This work bridges the gap between causal generation and global semantic alignment, demonstrating that even strictly autoregressive models can plan ahead.

Ensuring the safety of Large Language Models (LLMs) is a critical alignment challenge. Existing approaches often rely on invasive fine- tuning or external generation-based checks, which can be opaque and resource-inefficient. In this work, we investigate the geometry of safety concepts within pretrained representations, proposing a mechanistic methodology that identifies the layer where safe and unsafe concepts are maximally separable within a pretrained model’s representation space. By leveraging the intrinsic activation space of the optimal layer, we show that safety enforcement can be achieved via a simple linear classifier, avoiding the need for weight modification. We validate our framework across multiple domains (regulation, law, finance, cybersecurity, education, code, human resources, and social media), diverse tasks (safety classification, prompt injection, and toxicity detection), and 16 non-English languages on both encoder and decoder architectures. Our results show that: (i) the separation between safe and unsafe concepts emerges from a single layer direction in the activation space, (ii) monitoring internal representations provides a significantly more robust safeguarding mechanism compared to traditional evaluative or generative guardrail paradigms.

Despite the state-of-the-art performance of Large Language Models (LLMs) achieved on many tasks, their massive scale often leads to high computational and environmental costs, limiting their accessibility. Parameter-Efficient Fine-Tuning (PEFT) methods address this challenge by reducing the number of trainable parameters while maintaining strong downstream performance. Despite the advances in PEFT methods, current evaluations remain limited (in terms of evaluated models and datasets) and difficult to reproduce. To bridge this gap, we introduce PEFT-Bench, a unified end-to-end benchmark for evaluating diverse PEFT methods on autoregressive LLMs. We demonstrate its usage across 27 NLP datasets and 7 PEFT methods. To account for different PEFT training and inference factors, we also introduce the PEFT Soft Cost Penalties (PSCP) metric, which takes trainable parameters, inference speed, and training memory usage into account.

Accurately forecasting large stock price movements after corporate earnings announcements is a longstanding challenge. Existing methods–sentiment lexicons, fine-tuned encoders, and standalone LLMs–often **lack temporal-causal reasoning** and are prone to **narrative bias**, echoing overly optimistic managerial tone. We introduce **Context-Enriched Agentic RAG (CARAG)**, a retrieval-augmented framework that deploys a team of cooperative LLM agents, each specializing in a distinct analytical task: evaluating historical performance, assessing the credibility of guidance, or benchmarking against peers.Agents retrieve structured evidence from a Causal-Temporal Knowledge Graph (CTKG) built from financial statements and earnings calls, enabling grounded, context-rich reasoning. This design mitigates LLM hallucinations and produces more objective predictions.Without task-specific training, our system achieves state-of-the-art zero-shot performance across NASDAQ, NYSE, and MAEC datasets, outperforming both larger LLMs and fine-tuned models in macro-F1, MCC, and Sharpe, beating market benchmarks (S P 500 and Nasdaq) for the same forecasting horizon. Code, datasets, prompts, and implementation details are included in the supplementary material to ensure full reproducibility.

Automated Essay Scoring (AES) has gained increasing attention in recent years, yet research on Arabic AES remains limited due to the lack of publicly available datasets. To address this, we introduce LAILA, the largest publicly available Arabic AES dataset to date, comprising 7,859 essays annotated with holistic and trait-specific scores on seven dimensions: relevance, organization, vocabulary, style, development, mechanics, and grammar. We detail the dataset design, collection, and annotations, and provide benchmark results using state-of-the-art Arabic and English models in prompt-specific and cross-prompt settings. LAILA fills a critical need in Arabic AES research, supporting the development of robust scoring systems.

Large language models (LLMs) increasingly rely on external tools and APIs to execute complex tasks specified in natural language. Evaluating such tool-calling capabilities in realistic enterprise settings is challenging: APIs are often proprietary, heterogeneous, and difficult to share, limiting reproducible benchmarks. To address this, we introduce Live API Bench, a comprehensive benchmark constructed by transforming NL2SQL datasets into interactive API environments. Our pipeline converts SQL queries from BIRD-SQL into executable API sequences across three formulations—SLOT, SEL, and REST—covering minimal general-purpose operations, domain-specific multi-step tasks, and function-oriented RESTful interactions, respectively. The benchmark spans 11 databases with over 2,500 invocable tools, paired with human-authored queries, ground-truth API sequences, and verified final answers. Live API Bench enables systematic evaluation of core challenges in tool use, including error handling, sequential reasoning, parameter generation, response parsing, and robustness across diverse domains. We evaluate 10 LLMs and 4 ReACT agents, observing low task completion rates (7–47%), which improve modestly to 50% under interactive agent settings, highlighting substantial scope for improving LLM tool-calling performance. We release all code and data associated with this paper.

The intentional creation and spread of disinformation poses a significant threat to public discourse. However, existing English datasets and research rarely address the intentionality behind the disinformation. This work presents MALINT, the first human-annotated English corpus developed in collaboration with expert fact-checkers to capture disinformation and its malicious intent. We utilize our novel corpus to benchmark 12 language models, including small language models (SLMs) such as BERT and large language models (LLMs) like Llama 3.3, on binary and multilabel intent classification tasks. Moreover, inspired by inoculation theory from psychology and communication studies, we investigate whether incorporating knowledge of malicious intent can improve disinformation detection. To this end, we propose intent-based inoculation, an intent-augmented reasoning for LLMs that integrates intent analysis to mitigate the persuasive impact of disinformation. Analysis on six disinformation datasets, five LLMs, and seven languages shows that intent-augmented reasoning improves zero-shot disinformation detection. To support research in intent-aware disinformation detection, we release the MALINT dataset with annotations from each annotation step.

Training Large Language Models (LLMs) with high multilingual coverage is becoming increasingly important — especially when monolingual resources are scarce. Recent studies have found that LLMs process multilingual inputs in shared concept spaces, thought to support generalization and cross-lingual transfer. However, these prior studies often do not use causal methods, lack deeper error analysis or focus on the final model only, leaving open how these spaces emerge *during training*. We investigate the development of language-agnostic concept spaces during pretraining of EuroLLM through the causal interpretability method of activation patching. We isolate cross-lingual concept representations, then inject them into a translation prompt to investigate how consistently translations can be altered, independently of the language. We find that *shared concept spaces emerge early and continue to refine*, but that *alignment with them is language-dependent*. Furthermore, in contrast to prior work, our fine-grained manual analysis reveals that some apparent gains in translation quality reflect shifts in behavior — like selecting senses for polysemous words or translating instead of copying cross-lingual homographs — rather than improved translation ability. Our findings offer new insight into the training dynamics of cross-lingual alignment and the conditions under which causal interpretability methods offer meaningful insights in multilingual contexts.

Does the prior knowledge of the vision encoder constrain the capability boundary of Multi-modal Large Language Models (MLLMs)? While most existing research treats MLLMs as unified systems optimized through end-to-end training, the impact of vision encoder’s prior knowledge is seldom investigated. In this work, we introduce a novel metric Rank_e to quantify the effect of prior knowledge of the vision encoder on MLLM performance. Our analysis reveals a positive correlation between prior knowledge and MLLM performance. Moreover, we find that domain-specific fine-tuning using solely end-to-end visual question answering (VQA) data is insufficient, particularly for entities with low inherent visual prior knowledge. To address this issue, we propose VisPRE (Vision Prior Remediation), a two-stage training framework that explicitly incorporates prior knowledge at the vision encoder level. Experimental results demonstrate that augmenting vision encoder’s prior knowledge substantially boosts the visual understanding capabilities of MLLMs, offering a novel and effective strategy for improving performance, especially in scenarios involving uncommon visual entities.

Retrieval-augmented generation (RAG) is a prevalent approach for building LLM-based question-answering systems that can take advantage of external knowledge databases. Due to the complexity of real-world RAG systems, there are many potential causes for erroneous outputs. Understanding the range of errors that can occur in practice is crucial for robust deployment. We present a new taxonomy of the error types that can occur in realistic RAG systems, examples of each, and practical advice for addressing them. Additionally, we curate a dataset of erroneous RAG responses annotated by error types. We then propose an auto-evaluation method aligned with our taxonomy that can be used in practice to track and address errors during development. Code and data are available at https://github.com/layer6ai-labs/rag-error-classification.

In the global drive toward carbon neutrality, deeply coordinated smart energy systems underpin industrial transformation, yet their interdisciplinary, fragmented, and fast-evolving expertise prevents general-purpose LLMs, lacking domain knowledge and physical-constraint awareness, from delivering precise engineering-aligned inference and generation. To address these challenges, we introduce Helios, the first large language model tailored to the smart energy domain, together with a comprehensive suite of resources to advance LLM research in this field. Specifically, we develop Enersys, a multi-agent collaborative framework for end-to-end dataset construction, through which we produce: (1) the first smart energy knowledge base, EnerBase, to enrich the model’s foundational expertise; (2) the first instruction fine-tuning dataset, EnerInsruct, to strengthen performance on domain-specific downstream tasks; and (3) the first RLHF dataset, EnerReinforce, to align the model with human preferences and industry standards. Leveraging these resources, Helios undergoes large-scale pretraining, SFT, and RLHF. We also release EnerBench, the first benchmark for evaluating LLMs in smart energy scenarios, and demonstrate that our approach significantly enhances domain knowledge mastery, task execution accuracy, and alignment with human preferences.

Sparse Autoencoders (SAEs) are powerful tools for interpreting neural representations, yet their use in audio remains underexplored. We train SAEs across all encoder layers of Whisper and HuBERT, provide an extensive evaluation of their stability, interpretability, and show their practical utility. Over 50% of the features remain consistent across random seeds, and reconstruction quality is preserved. SAE features capture general acoustic and semantic information as well as specific events, including environmental noises and paralinguistic sounds (e.g. laughter, whispering) and disentangle them effectively, requiring removal of only 19-27% of features to erase a concept. Feature steering reduces Whisper’s false speech detections by 70% with negligible WER increase, demonstrating real-world applicability. Finally, we find SAE features correlated with human EEG activity during speech perception, indicating alignment with human neural processing. The code and checkpoints are available at https://github.com/audiosae/audiosae_demo.

Recent studies suggest that transformer-based vision-language models (VLMs) capture the multimodality of concept processing in the human brain. However, a systematic evaluation exploring different types of VLM architectures and the role played by visual and textual context is still lacking. Here, we analyse multiple VLMs employing different strategies to integrate visual and textual modalities, along with language-only counterparts. We measure the alignment between concept representations by models and existing (fMRI) brain responses to concept words presented in two experimental conditions, where either visual (pictures) or textual (sentences) context is provided. Our results reveal that VLMs outperform the language-only counterparts in both experimental conditions. However, controlled ablation studies show that only for some VLMs, such as LXMERT and IDEFICS2, brain alignment stems from genuinely learning more human-like concepts during _pretraining_, while others are highly sensitive to the context provided at _inference_. Additionally, we find that vision-language encoders are more brain-aligned than more recent, generative VLMs. Altogether, our study shows that VLMs align with human neural representations in concept processing, while highlighting differences among architectures. We open-source code and materials to reproduce our experiments at: [https://github.com/dmg-illc/vl-concept-processing](https://github.com/dmg-illc/vl-concept-processing).

The growing collaboration between humans and AI models in generative tasks has introduced new challenges in distinguishing between *human-written*, *LLM-generated*, and *human-LLM collaborative* texts. In this work, we collect a multilingual, multi-domain, multi-generator dataset *FAIDSet*. We further introduce a fine-grained detection framework *FAID* to classify text into these three categories, and also to identify the underlying LLM family of the generator. Unlike existing binary classifiers, FAID is built to capture both authorship and model-specific characteristics. Our method combines multi-level contrastive learning with multi-task auxiliary classification to learn subtle stylistic cues. By modeling LLM families as distinct stylistic entities, we incorporate an adaptation to address distributional shifts without retraining for unseen data. Our experimental results demonstrate that FAID outperforms several baselines, particularly enhancing the generalization accuracy on unseen domains and new LLMs, thus offering a potential solution for improving transparency and accountability in AI-assisted writing. Our data and code are available at [https://github.com/mbzuai-nlp/FAID](https://github.com/mbzuai-nlp/FAID)

We present BabyBabelLM, a multilingual collection of datasets modeling the language a person observes from birth until they acquire a native language. We curate developmentally plausible pretraining data aiming to cover the equivalent of 100M English words of content in each of 45 languages. We compile evaluation suites and train baseline models in each language. BabyBabelLM aims to facilitate multilingual pretraining and cognitive modeling.

Large Language Models (LLMs) are integral to applications such as conversational agents and content creation, where precise control over a model’s personality is essential for maintaining tone, consistency, and user engagement. However, prevailing prompt-based or fine-tuning approaches either lack robustness or demand large-scale training data, making them costly and impractical. In this paper, we present PALETTE (Personality Adjustment by LLM SElf-TargeTed quEries), a novel method for personality editing in LLMs. Our approach introduces adjustment queries, where self-referential statements grounded in psychological constructs are treated analogously to factual knowledge, enabling direct editing of personality-related responses. Unlike fine-tuning, PALETTE requires only 12 editing samples to achieve substantial improvements in personality alignment across personality dimensions. Experimental results from both automatic and human evaluations demonstrate that our method enables more stable and well-balanced personality control in LLMs.

Large language models (LLMs) are increasingly adopted for handling structured data, including tabular and relational inputs, despite mostly being pretrained on unstructured text. This raises a key question: how effectively do pretrained representations from language-focused LLMs transfer to tasks involving structured inputs? We address this through controlled experiments using two small open-source LLMs, systematically re-initializing subsets of layers with random weights before fine-tuning on structured datasets and comparing results to unstructured datasets. Our analyses show that, for structured data, most pretrained depth contributes little, with performance often saturating after the first few layers, whereas unstructured tasks benefit more consistently from deeper pretrained representations. Pretraining remains useful mainly in low-resource settings, with its impact diminishing as more training data becomes available.

Despite their impressive performance, vision-language models (VLMs) still struggle on culturally situated inputs. To understand how VLMs process culturally grounded information, we study the presence of culture-sensitive neurons, i.e., neurons whose activations show preferential sensitivity to inputs associated with particular cultural contexts. We examine whether such neurons are important for culturally diverse visual question answering and where they are located. Using the CVQA benchmark, we identify neurons of culture selectivity and perform diagnostic tests by deactivating the neurons flagged by various identification methods. Experiments on three VLMs across 25 cultural groups demonstrate the existence of neurons whose ablation disproportionately harms performance on questions about the corresponding cultures, while having limited effects on others. Moreover, we introduce a new margin-based selector—Contrastive Activation Margin (ConAct)—and show that it outperforms probability- and entropy-based methods in identifying neurons associated with cultural selectivity. Finally, our layer-wise analyses reveal that such neurons are not uniformly distributed: they cluster in specific decoder layers in a model-dependent way.

Multiple-Choice Questions (MCQs) are often used to assess knowledge, reasoning abilities, and even values encoded in large language models (LLMs). While the effect of multilingualism has been studied on LLM factual recall, this paper seeks to investigate the less explored question of language-induced variation in value-laden MCQ responses. Are multilingual LLMs consistent in their responses across languages, *i.e.* behave like theoretical *polyglots*, or do they answer value-laden MCQs depending on the language of the question, like a *multitude* of monolingual models expressing different values through a single model? We release a new corpus, the Multilingual European Value Survey (**MEVS**), which, unlike prior work relying on machine translation or ad hoc prompts, solely comprises human-translated survey questions aligned in 8 European languages. We administer a subset of those questions to over thirty multilingual LLMs of various sizes, manufacturers and alignment-fine-tuning status under comprehensive, controlled prompt variations including answer order, symbol type, and tail character. Our results show that while larger, instruction-tuned models display higher overall consistency, the robustness of their responses varies greatly across questions, with certain MCQs eliciting total agreement *within and across* models while others leave LLM answers split. Language-specific behavior seems to arise in all consistent, instruction-fine-tuned models, but only on certain questions, warranting a further study of the selective effect of preference fine-tuning.

Understanding fine-grained links between documents is crucial for many applications, yet progress is limited by the lack of efficient methods for for data curation. To address this limitation, we introduce a domain-agnostic framework for bootstrapping sentence-level cross-document links from scratch. Our approach (1) generates and validates semi-synthetic datasets of linked documents, (2) uses these datasets to benchmark and shortlist the best-performing linking approaches, and (3) applies the shortlisted methods in large-scale human-in-the-loop annotation of natural text pairs. We apply the framework in two distinct domains – peer review and news – and show that combining retrieval models with LLMs achieves a 73% human approval rate for suggested links, more than doubling the acceptance of strong retrievers alone. Our framework allows users to produce novel datasets that enable systematic study of cross-document understanding, supporting downstream tasks such as media framing analysis and peer review assessment. All code, data, and annotation protocols are released to facilitate future research.

Meta-reviewing is a pivotal stage in the peer-review process, serving as the final step in determining whether a paper is recommended for acceptance. Prior research on meta-reviewing has treated this as a summarization problem over review reports. However, complementary to this perspective, meta-reviewing is a decision-making process that requires weighing reviewer arguments and placing them within a broader context. Prior research has demonstrated that decision-makers can be effectively assisted in such scenarios via dialogue agents. In line with this framing, we explore the practical challenges for realizing dialog agents that can effectively assist meta-reviewers. Concretely, we first address the issue of data scarcity for training dialogue agents by generating synthetic data using Large Language Models (LLMs) based on a self-refinement strategy to improve the relevance of these dialogues to expert domains. Our experiments demonstrate that this method produces higher-quality synthetic data and can serve as a valuable resource towards training meta-reviewing assistants. Subsequently, we utilize this data to train dialogue agents tailored for meta-reviewing and find that these agents outperform *off-the-shelf* LLM-based assistants for this task. Finally, we apply our agents in real-world meta-reviewing scenarios and confirm their effectiveness in enhancing the efficiency of meta-reviewing.

The factual reliability of Large Language Models (LLMs) remains a critical barrier to their adoption in high-stakes domains due to their propensity to hallucinate. Current detection methods often rely on surface-level signals from the model’s output, overlooking the failures that occur within the model’s internal reasoning process. In this paper, we introduce a new paradigm for hallucination detection by analyzing the dynamic topology of the evolution of model’s layer-wise attention. We model the sequence of attention matrices as a zigzag graph filtration and use zigzag persistence, a tool from Topological Data Analysis, to extract a topological signature. Our core hypothesis is that factual and hallucinated generations exhibit distinct topological signatures. We validate our framework, HalluZig, on multiple benchmarks, demonstrating that it outperforms strong baselines. Furthermore, our analysis reveals that these topological signatures are generalizable across different models and can enable early detection of hallucinations.

Large language models (LLMs) have demonstrated strong performance in a wide-range of language tasks without requiring task-specific fine-tuning. However, they remain prone to hallucinations and inconsistencies, and often struggle with complex reasoning, in part due to the limitations of autoregressive generation. We propose to address some of these issues, particularly for structured prediction, by combining LLMs with combinatorial inference to marry the predictive power of LLMs with the structural consistency provided by inference methods. We perform exhaustive experiments in an effort to understand which prompting strategies can best estimate confidence values for downstream symbolic inference, and find that, independent of prompting strategy, incorporating symbolic inference yields more consistent and accurate predictions than prompting alone. Finally, we show that calibration and fine-tuning with structured learning objectives further increases performance on challenging tasks, highlighting that structured learning remains valuable in the era of LLMs.

Large Language Models and Vision-Language Models have achieved impressive performance across a wide range of tasks, yet they remain vulnerable to carefully crafted perturbations. In this study, we seek to pinpoint the sources of this fragility by identifying parameters and input dimensions (pixels or token embeddings) that are susceptible to such perturbations. To this end, we propose a stability measure called FI, First order local Influence, which is rooted in information geometry and quantifies the sensitivity of individual parameter and input dimensions. Our extensive analysis across LLMs and VLMs (from 1.5B to 13B parameters) reveals that: (I) A small subset of parameters or input dimensions with high FI values disproportionately contribute to model brittleness. (II) Mitigating the influence of these vulnerable parameters during model merging leads to improved performance.

Standard autoregressive decoding in large language models (LLMs) is inherently short-sighted, often failing to find globally optimal reasoning paths due to its token-by-token generation process. While inference-time strategies like foresight sampling attempt to mitigate this by simulating future steps, they typically rely on ad-hoc heuristics for valuing paths and pruning the search space. This paper introduces Martingale Foresight Sampling (MFS), a principled framework that reformulates LLM decoding as a problem of identifying an optimal stochastic process. By modeling the quality of a reasoning path as a stochastic process, we leverage Martingale theory to design a theoretically-grounded algorithm. Our approach replaces heuristic mechanisms with principles from probability theory: step valuation is derived from the Doob Decomposition Theorem to measure a path’s predictable advantage, path selection uses Optional Stopping Theory for principled pruning of suboptimal candidates, and an adaptive stopping rule based on the Martingale Convergence Theorem terminates exploration once a path’s quality has provably converged. Experiments on six reasoning benchmarks demonstrate that MFS surpasses state-of-the-art methods in accuracy while significantly improving computational efficiency. Code will be released at https://github.com/miraclehetech/EACL2026-Martingale-Foresight-Sampling.

The in-context learning (ICL) coding, reasoning, and tool-using ability of LLMs has spurred interest in library learning (i.e., the creation and exploitation of reusable and composable functions, tools, or lemmas). Such systems often promise improved task performance and computational efficiency by caching reasoning (i.e., storing generated tools) - all without finetuning. However, we find strong reasons to be skeptical. Specifically, we identify a serious evaluation flaw present in a large number of ICL library learning works: these works do not correct for the difference in computational cost between baseline and library learning systems. Studying three separately published ICL library learning systems, we find that all of them fail to consistently outperform the simple baseline of prompting the model - improvements in task accuracy often vanish or reverse once computational cost is accounted for. Furthermore, we perform an in-depth examination of one such system, LEGO-Prover, which purports to learn reusable lemmas for mathematical reasoning. We find no evidence of the direct reuse of learned lemmas, and find evidence against the soft reuse of learned lemmas (i.e., reuse by modifying relevant examples).Our findings suggest that a serious re-examination of the effectiveness of ICL LLM-based library learning is required, as is much stronger standards for evaluation. An equal computational budget must be used for baselines, alongside behavioural analysis.

Long-form videos that span across wide temporal intervals are highly information redundant and contain multiple distinct events or entities that are often loosely related. Therefore, when performing long-form video question answering (LVQA), all information necessary to generate a correct response can often be contained within a small subset of frames. Recent literature leverage large language models (LLMs) in LVQA benchmarks, achieving exceptional performance, while relying on vision language models (VLMs) to convert all visual content within videos into natural language. Such VLMs often independently caption a large number of frames uniformly sampled from long videos, which is not efficient and can mostly be redundant. Motivated by this inefficiency, we propose LVNet, a modular and training-free framework featuring a novel Hierarchical Keyframe Selector (HKS) that efficiently selects a minimal set of informative frames tailored to each question. LVNet’s modularity allows easy integration with existing approaches for more efficient LVQA. We achieve state-of-the-art performance among similarly configured models across four benchmark LVQA datasets: EgoSchema, NExT-QA, IntentQA, VideoMME. The code can be found athttps://github.com/jongwoopark7978/LVNet

Interest in leveraging Large Language Models (LLMs) for emotional support systems motivates the need to understand how these models comprehend and represent emotions internally. While recent works show the presence of emotion concepts in the hidden state representations, it’s unclear if the model has a robust representation that is consistent across different datasets. In this paper, we present a unified view to understand emotion representation in LLMs, experimenting with diverse datasets and prompts. We then evaluate the reasoning ability of the models on a complex emotion identification task. We find that LLMs have a common emotion representation in the later layers of the model, and the vectors capturing the direction of emotions extracted from these representations can be interchanged among datasets with minimal impact on performance. Our analysis of reasoning with Chain of Thought (CoT) prompting shows the limits of emotion comprehension. Therefore, despite LLMs implicitly having emotion representations, they are not equally skilled at reasoning with them in complex scenarios. This motivates the need for further research to find new approaches.

Reliable mathematical and scientific reasoning remains an open challenge for large vision–language models (VLMs). Standard final-answer evaluation often masks reasoning errors, allowing silent failures to persist. To address this gap, we introduce TRACE (Transparent Reasoning And Consistency Evaluation), a framework for analyzing, diagnosing, and improving reasoning in VLMs. At its core, TRACE leverages Auxiliary Reasoning Sets (ARS), compact sub-question–answer pairs that decompose complex problems, evaluate intermediate steps through consistency-based metrics, and expose failures overlooked by standard evaluation. Our experiments show that consistency across ARS is linked to final-answer correctness and helps pinpoint the reasoning steps where failures arise, offering actionable signals for model improvement.

We introduce Argument Representation Coverage (ARC), a bottom-up evaluation framework that assesses how well summaries preserve structured salient arguments, a crucial issue in summarizing high-stakes domains such as law. ARC provides an interpretable lens by distinguishing between different information types to be covered and by separating omissions from factual errors.Using ARC, we evaluate summaries from eight open-weight LLMs in two domains where argument roles are central: long legal opinions and scientific articles. Our results show that while LLMs capture some salient roles, they frequently omit critical information, particularly when arguments are sparsely distributed across the input. Moreover, ARC uncovers systematic patterns—showing how context window positional bias and role-specific preferences shape argument coverage—providing actionable guidance for developing more complete and reliable summarization strategies.

Current speech evaluation suffers from two critical limitations: the need and difficulty of designing specialized systems targeting individual audio characteristics, and poor correlation between automatic evaluation methods and human preferences. This work presents a systematic study of Large Audio Model (LAM) as a Judge, AudioJudge, investigating whether it can provide a unified evaluation framework that addresses both challenges. We systematically explore AudioJudge across audio characteristic detection tasks, including pronunciation, speaking rate, speaker identification and speech quality, and system-level human preference simulation for automated benchmarking. We investigate different prompt engineering strategies, finding that audio concatenation combined with in-context learning significantly improves performance across both audio characteristic detection and human preference simulation tasks. We further introduce a multi-aspect ensemble AudioJudge to enable general-purpose multi-aspect audio evaluation. This method decomposes speech assessment into specialized judges for lexical content, speech quality, and paralinguistic features, achieving up to 0.91 Spearman correlation with human preferences on our system ranking benchmark. Robustness analysis reveals that while LAMs maintain strong performance under acoustic noise, they exhibit significant verbosity and positional biases that require careful mitigation.

Large Language Models (LLMs) are highly effective at adapting to users’ styles, preferences, and contextual signals—a property that underlies much of their practical usefulness, but which can even manifest as sycophancy, i.e., alignment with user-implied beliefs evenwhen these contradict factual accuracy or rational reasoning. Prior work treats sycophancy as a surface-level artefact addressed via inference-time or post-hoc methods. We argue that it is a policy-level failure arising from missing agentic control over agreement under pressure. To make sycophancy amenable to explicit control, we propose learning agentic policies modelling LLMs’ behaviour as a decision-making problem. Our approach equips a single model with an explicit action space that includes answering directly, countering misleading signals, or asking for clarification. The policy is trained to optimise a multi-objective reward that balances task success, sycophancy resistance, and behavioural consistency via a control mechanism that operates through agentic behaviour. We evaluate the method on different benchmarks, showing that the approaches reduce sycophancy, improving performance, and generalise robustly across languages. These findings suggest that mitigating sycophancy requires moving beyond compliance-oriented generation towards agreement-agentic control.

While large language models (LLMs) offer great promise, they also pose concrete safety risks. To audit and mitigate these risks, researchers have developed automated red-teaming methods, which generate adversarial prompts to elicit unsafe behavior of target LLMs during evaluation. Recent automated red-teaming methods for LLMs face a persistent trade-off: techniques that increase prompt diversity often reduce the level of the toxicity elicited from the target LLMs, while toxicity-maximizing methods tend to collapse diversity. To address the limitations, we propose ToxiPrompt, a two-stage framework that explicitly separates exploration (diversity) from exploitation (toxicity) and reunifies them with a single selection criterion to balance between diversity and toxicity. Experimental results show that ToxiPrompt outperforms four state-of-the-art baselines in both adversarial prompt diversity and the level of elicited toxicity from target LLMs, improving 14.6% harmonic mean of toxicity and diversity against the best baseline. The approach also performs well for multiple instruction-tuned target LLMs (Llama-2/3, Qwen, Mistral) without re-tuning, achieving up to 55% harmonic mean improvement against the best baseline. Our code is available at https://github.com/seungho715/ToxiPrompt

Text embeddings are an essential building component of several NLP tasks such as retrieval-augmented generation which is crucial for preventing hallucinations in LLMs. Despite the recent release of massively multilingual MTEB (MMTEB), African languages remain underrepresented, with existing tasks often repurposed from translation benchmarks such as FLORES clustering or SIB-200. In this paper, we introduce AfriMTEB—a regional expansion of MMTEB covering 59 languages, 14 tasks, and 38 datasets, including six newly added datasets. Unlike many MMTEB datasets that include fewer than five languages, the new additions span 14 to 56 African languages and introduce entirely new tasks, such as hate speech detection, intent detection, and emotion classification, which were not previously covered. Complementing this, we present AfriE5, an adaptation of the instruction-tuned mE5 model to African languages through cross-lingual contrastive distillation. Our evaluation shows that AfriE5 achieves state-of-the-art performance, outperforming strong baselines such as Gemini-Embeddings and mE5.

Generalization to unseen concepts is a central challenge due to the scarcity of human annotations in Mention-agnostic Biomedical Concept Recognition (MA-BCR). This work makes two key contributions to systematically address this issue. First, we propose an evaluation framework built on hierarchical concept indices and novel metrics to measure generalization. Second, we explore LLM-based Auto-Labeled Data (ALD) as a scalable resource, creating a task-specific pipeline for its generation. Our research unequivocally shows that while LLM-generated ALD cannot fully substitute for manual annotations, it is a valuable resource for improving generalization, successfully providing models with the broader coverage and structural knowledge needed to approach recognizing unseen concepts. Code and datasets are available at https://github.com/bio-ie-tool/hi-ald.

Zero-shot dense retrieval is a challenging setting where a document corpus is provided without relevant queries, necessitating a reliance on pretrained dense retrievers (DRs). However, since these DRs are not trained on the target corpus, they struggle to represent semantic differences between similar documents. To address this failing, we introduce a training-free representation sharpening framework that augments a document’s representation with information that helps differentiate it from similar documents in the corpus. On over twenty datasets spanning multiple languages, the representation sharpening framework proves consistently superior to traditional retrieval, setting a new state-of-the-art on the BRIGHT benchmark. We show that representation sharpening is compatible with prior approaches to zero-shot dense retrieval and consistently improves their performance. Finally, we address the performance-cost tradeoff presented by our framework and devise an indexing-time approximation that preserves the majority of our performance gains over traditional retrieval, yet suffers no additional inference-time cost.

In many real-world applications, users rely on natural language instructions to guide large language models (LLMs) across a wide range of tasks. These instructions are often complex, diverse, and subject to frequent change. However, LLMs do not always attend to these instructions reliably, and users lack simple mechanisms to emphasize their importance beyond modifying prompt wording or structure. To address this, we present an inference-time method that enables users to emphasize specific parts of their prompt by steering the model’s attention toward them, aligning the model’s perceived importance of different prompt tokens with user intent. Unlike prior approaches that are limited to static instructions, require significant offline profiling, or rely on fixed biases, we dynamically update the proportion of model attention given to the user-specified parts–ensuring improved instruction following without performance degradation. We demonstrate that our approach improves instruction following across a variety of tasks involving multiple instructions and generalizes across models of varying scales.

End-to-end form filling refers to automatically populating fields in a document-style form with the appropriate information derived from external data. Although prevalent and useful, no formal benchmark exists for evaluating systems’ form completion accuracy. Existing datasets focus on parsing, extraction and web form interaction, rather than end-to-end completion of document-style forms. We propose FormGym, a benchmark formulation of the end-to-end form filling task that evaluates form completion and accuracy. We construct FormGym by repurposing three existing datasets and add one new dataset to achieve more challenging, diverse, and realistic test cases. Our studies show baseline vision language agents (VLAs) perform poorly on FormGym in every scenario, primarily due to poor field localization. GUI agents perform better but suffer from high latency and costs. Therefore we also introduce FieldFinder, a field localization tool that enables zero-shot VLAs to find and accurately place text in input fields. We find that VLAs augmented with FieldFinder achieve better performance compared to baselines in all models.

We present NarraBench, a theory-informed taxonomy of narrative-understanding tasks, as well as an associated survey of 78 existing benchmarks in the area. We find significant need for new evaluations covering aspects of narrative understanding that are either overlooked in current work or are poorly aligned with existing metrics. Specifically, we estimate that only 27% of narrative tasks are well captured by existing benchmarks, and we note that some areas – including narrative events, style, perspective, and revelation – are nearly absent from current evaluations. We also note the need for increased development of benchmarks capable of assessing constitutively subjective and perspectival aspects of narrative, that is, aspects for which there is generally no single correct answer. Our taxonomy, survey, and methodology are of value to NLP researchers seeking to test LLM narrative understanding.

Large language models (LLMs) increasingly produce natural language explanations, yet these explanations often lack faithfulness, and they do not reliably reflect the evidence the model uses to decide. We introduce FaithLM, a model-agnostic framework that evaluates and improves the faithfulness of LLM explanations without token masking or task-specific heuristics. FaithLM formalizes explanation faithfulness as an intervention property: a faithful explanation should yield a prediction shift when its content is contradicted. Theoretical analysis shows that the resulting contrary-hint score is a sound and discriminative estimator of faithfulness. Building on this principle, FaithLM iteratively refines both the elicitation prompt and the explanation to maximize the measured score. Experiments on three multi-domain datasets and multiple LLM backbones demonstrate that FaithLM consistently increases faithfulness and produces explanations more aligned with human rationales than strong self-explanation baselines. These findings highlight that intervention-based evaluation, coupled with iterative optimization, provides a principled route toward faithful and reliable LLM explanations.

The Uniform Information Density (UID) hypothesis posits that speakers are subject to a communicative pressure to distribute information evenly within utterances, minimising surprisal variance. While this hypothesis has been tested empirically, prior studies are limited exclusively to text-only inputs, abstracting away from the perceptual context in which utterances are produced. In this work, we present the first computational study of UID in visually grounded settings. We estimate surprisal using multilingual vision-and-language models over image–caption data in 30 languages and visual storytelling data in 13 languages, together spanning 11 families. We find that grounding on perception consistently smooths the distribution of information, increasing both global and local uniformity across typologically diverse languages compared to text-only settings. In visual narratives, grounding in both image and discourse contexts has additional effects, with the strongest surprisal reductions occurring at the onset of discourse units. Overall, this study takes a first step towards modelling the temporal dynamics of information flow in ecologically plausible, multimodal language use, and finds that grounded language exhibits greater information uniformity, supporting a context-sensitive formulation of UID.

Several previous works concluded that the largest part of generation capabilities of large language models (LLM) are learned (early) during pre-training. However, LLMs still require further alignment to adhere to downstream task requirements and stylistic preferences, among other desired properties. As LLMs continue to scale in terms of size, the computational cost of alignment procedures increase prohibitively.In this work, we propose a novel approach to circumvent these costs via proxy-based test-time alignment, i.e. using guidance from a small aligned model. Our approach can be described as a token-specific cascading method, where the token-specific deferral rule is reduced to 0-1 knapsack problem. In this setting, we derive primal and dual approximations of the optimal deferral decision. We experimentally show the benefits of our method both in task performance and speculative decoding speed.

Evaluating Large Language Models (LLMs) for mental health support poses unique challenges to reliable evaluation due to the emotionally and cognitively complex nature of therapeutic dialogue. Existing benchmarks are limited in scale, authenticity, and reliability, often relying on synthetic or social media data, and lack frameworks to assess when automated judges can be trusted. To address the need for large-scale authentic dialogue datasets and judge-reliability assessment, we introduce two benchmarks that provide a framework for generation and evaluation in this domain. MentalBench-100k consolidates 10,000 authentic single-session therapeutic conversations from three real-world scenarios datasets, each paired with nine LLM-generated responses, yielding 100,000 response pairs. MentalAlign-70k reframes evaluation by comparing four high-performing LLM judges with human experts across 70,000 ratings on seven attributes, grouped into Cognitive Support Score (CSS) and Affective Resonance Score (ARS). We then employ the Affective–Cognitive Agreement Framework, a statistical methodology using intraclass correlation coefficients (ICC) with confidence intervals to quantify agreement, consistency, and bias between LLM judges and human experts. Our analysis reveals systematic inflation by LLM judges, strong reliability for cognitive attributes such as guidance and informativeness, reduced precision for empathy, and some unreliability in safety and relevance. Our contributions establish new methodological and empirical foundations for the reliable and large-scale evaluation of LLMs in mental health contexts.

We propose a novel self-attention mechanism for document understanding that takes into account text block positions in relative polar coordinate system rather than the Cartesian one. Based on this mechanism, we build DocPolarBERT, a layout-aware BERT model for document understanding that eliminates the need for absolute 2D positional embeddings. Despite being pre-trained on a dataset more than six times smaller than the widely used IIT-CDIP corpus, DocPolarBERT achieves state-of-the-art results. These results demonstrate that a carefully designed attention mechanism can compensate for reduced pre-training data, offering an efficient and effective alternative for document understanding.

Large language models (LLMs) are increasingly used to produce open-ended, interpretive annotations, yet there is no validated, scalable measure of ***idea-level similarity*** to expert annotations. We (i) introduce the content evaluation of LLM annotations as a core, understudied task, (ii) propose IDEAlign for capturing expert similarity judgments via the odd-one-out tasks, and (iii) benchmark various similarity methods, such as text embeddings, topic models, and LLM-as-a-judge, against these human ratings. Applying this approach to two real-world educational datasets (interpreting math reasoning and feedback generation), we find that most metrics fail to capture the nuanced dimensions of similarity meaningful to experts. LLM-as-a-judge performs best (11–18% improvement over other methods) but still falls short of expert alignment, making it useful as a triage filter rather than a substitute for human review. Our work demonstrates the difficulty of evaluating open-ended LLM annotations at scale, and positions IDEAlign as a reusable protocol for benchmarking on this task, thereby informing responsible deployment of LLMs.

Long-term conversational agents face a fundamental scalability challenge as interactions extend over time: repeatedly processing entire conversation histories becomes computationally prohibitive. Current approaches attempt to solve this through memory frameworks that predominantly fragment conversations into isolated embeddings or graph representations and retrieve relevant ones in a RAG style. While computationally efficient, these methods often treat memory formation minimally and fail to capture the subtlety and coherence of human memory. We introduce Amory, a working memory framework that actively constructs structured memory representations through enhancing agentic reasoning during offline time. Amory organizes conversational fragments into episodic narratives, consolidates memories with momentum, and semanticizes peripheral facts into semantic memory. At retrieval time, the system employs coherence-driven reasoning over narrative structures. Evaluated on the LOCOMO benchmark for long-term reasoning, Amory achieves considerable improvements over previous state-of-the-art, with performance comparable to full context reasoning while reducing response time by 50%. Analysis shows that momentum-aware consolidation significantly enhances response quality, while coherence-driven retrieval provides superior memory coverage compared to embedding-based approaches.

Despite the recent advancements in NLP with the advent of Large Language Models (LLMs), Entity Linking (EL) for historical texts remains challenging due to linguistic variation, noisy inputs, and evolving semantic conventions. Existing solutions either require substantial training data or rely on domain-specific rules that limit scalability. In this paper, we present MHEL-LLaMo (Multilingual Historical Entity Linking with Large Language MOdels), an unsupervised ensemble approach combining a Small Language Model (SLM) and an LLM. MHEL-LLaMo leverages a multilingual bi-encoder (BELA) for candidate retrieval and an instruction-tuned LLM for NIL prediction and candidate selection via prompt chaining. Our system uses SLM’s confidence scores to discriminate between easy and hard samples, applying an LLM only for hard cases. This strategy reduces computational costs while preventing hallucinations on straightforward cases. We evaluate MHEL-LLaMo on four established benchmarks in six European languages (English, Finnish, French, German, Italian and Swedish) from the 19th and 20th centuries. Results demonstrate that MHEL-LLaMo outperforms state-of-the-art models without requiring fine-tuning, offering a scalable solution for low-resource historical EL. Our error analysis reveals that 41% of false predictions exhibit semantic proximity to ground truth entities, highlighting the LLM’s accurate disambiguation of historical references.

Text-to-image (T2I) models are increasingly employed by users worldwide. However, prior research has pointed to the high sensitivity of T2I towards particular input languages - when faced with languages other than English (i.e., different surface forms of the same prompt), T2I models often produce culturally stereotypical depictions, prioritizing the surface over the prompt’s semantics. Yet a comprehensive analysis of this behavior, which we dub Surface-over-Semantics (SoS), is missing. We present the first analysis of T2I models’ SoS tendencies. To this end, we create a set of prompts covering 171 cultural identities, translated into 14 languages, and use it to prompt seven T2I models. To quantify SoS tendencies across models, languages, and cultures, we introduce a novel measure and analyze how the tendencies we identify manifest visually. We show that all but one model exhibit strong surface-level tendency in at least two languages, with this effect intensifying across the layers of T2I text encoders. Moreover, these surface tendencies frequently correlate with stereotypical visual depictions.

Politeness is an important social phenomenon which influences the flow of conversations. Several studies proposed models to discover and analyze linguistic cues associated with (im)polite language. However, no prior work computationally studied how politeness perception interacts with other social dimensions such as gender. We propose a model for automatic discovery of linguistic patterns which correlate with disagreement in politeness annotations, specifically focusing on gender differences. The model discovers fine-grained context patterns of words which correlate with disagreement in politeness annotations between men and women annotators. We apply the proposed model on emails annotated for politeness. Results show women rate emails which contain formal cues (e.g. To whom it may concern) more polite than men annotators rate them, while men rate emails exhibiting informal language cues (e.g. haven’t seen my new swing) more polite than women annotators rate them. Our findings highlight the importance of studying politeness through multiple demographic perspectives.

Time alters the visual appearance of entities in our world, like objects, places, and animals. Thus, for accurately generating contextually-relevant images, knowledge and reasoning about time can be crucial (e.g., for generating a landscape in spring vs. in winter). Yet, although substantial work exists on understanding and improving temporal knowledge in natural language processing, research on how temporal phenomena appear and are handled in text-to-image (T2I) models remains scarce. We address this gap with TempViz, the first data set to holistically evaluate temporal knowledge in image generation, consisting of 7.9k prompts and more than 600 reference images. Using TempViz, we study the capabilities of five T2I models across five temporal knowledge categories. Human evaluation shows that temporal competence is generally weak, with no model exceeding 75% accuracy across categories. Towards larger-scale studies, we also examine automated evaluation methods, comparing several established approaches against human judgments. However, none of these approaches provides a reliable assessment of temporal cues - further indicating the pressing need for future research on temporal knowledge in T2I.

Augmenting toxic language data in a controllable and class-specific manner is crucial for improving robustness in toxicity classification, yet remains challenging due to limited supervision and distributional skew. We propose ToxiGAN, a class-aware text augmentation framework that combines adversarial generation with semantic guidance from large language models (LLMs). To address common issues in GAN-based augmentation such as mode collapse and semantic drift, ToxiGAN introduces a two-step directional training strategy and leverages LLM-generated neutral texts as semantic ballast. Unlike prior work that treats LLMs as static generators, our approach dynamically selects neutral exemplars to provide balanced guidance. Toxic samples are explicitly optimized to diverge from these exemplars, reinforcing class-specific contrastive signals. Experiments on four hate speech benchmarks show that ToxiGAN achieves the strongest average performance in both macro-F1 and hate-F1, consistently outperforming traditional and LLM-based augmentation methods. Ablation and sensitivity analyses further confirm the benefits of semantic ballast and directional training in enhancing classifier robustness.

The basic underlying assumption of machine learning (ML) models is that the training and test data are sampled from the same distribution. However, in daily practice, this assumption is often broken, i.e. the distribution of the test data changes over time, which hinders the application of conventional ML models. One domain where the distribution shift naturally occurs is text classification, since people always find new topics to discuss. To this end, we survey research articles studying open-set text classification and related tasks. We divide the methods in this area based on the constraints that define the kind of distribution shift and the corresponding problem formulation, i.e. learning with the Universum, zero-shot learning, and open-set learning. We next discuss the predominant mitigation approaches for each problem setup. We further identify several future work directions, aiming to push the boundaries beyond the state of the art. Finally, we explain how continual learning can solve many of the issues caused by the shifting class distribution. We maintain a list of relevant papers at https://github.com/Eduard6421/Open-Set-Survey.

Reasoning has become a central paradigm for large language models (LLMs), consistently boosting accuracy across diverse benchmarks. Yet its suitability for precision-sensitive use remains unclear. We present the first systematic study of reasoning for classification tasks under strict low false positive rate (FPR) regimes. Our analysis covers two tasks—safety detection and hallucination detection—evaluated in both fine-tuned and zero-shot settings, using standard LLMs and Large Reasoning Models (LRMs). Our results reveal a clear trade-off: Think On (reasoning-augmented) generation improves overall accuracy, but performs poorly at the low-FPR thresholds essential for practical use. In contrast, Think Off (no reasoning during inference) dominates in these precision-sensitive regimes, with Think On surpassing only when higher FPRs are acceptable. In addition, we find token-based scoring substantially outperforms self-verbalized confidence for precision-sensitive deployments. Finally, a simple ensemble of the two modes recovers the strengths of each. Taken together, our findings position reasoning as a double-edged tool: beneficial for average accuracy, but often ill-suited for applications requiring strict precision.

Preparing high-quality instructional materials remains a labor-intensive process that often requires extensive coordination among teaching faculty, instructional designers, and teaching assistants. In this work, we present Instructional Agents, a multi-agent large language model (LLM) framework designed to automate end-to-end course material generation, including syllabus creation, lecture scripts, LaTeX-based slides, and assessments. Unlike existing AI-assisted educational tools that focus on isolated tasks, Instructional Agents simulates role-based collaboration among educational agents to produce cohesive and pedagogically aligned content. The system operates in four modes: Autonomous, Catalog-Guided, Feedback-Guided, and Full Co-Pilot mode, enabling flexible control over the degree of human involvement. We evaluate Instructional Agents across five university-level computer science courses and show that it produces high-quality instructional materials while significantly reducing development time and human workload. By supporting institutions with limited instructional design capacity, Instructional Agents provides a scalable and cost-effective framework to democratize access to high-quality education, particularly in underserved or resource-constrained settings.

Document understanding requires modeling both structural and semantic relationships between the layout elements within the document, with human-perceived reading order (RO) playing a crucial yet often neglected role compared to heuristic OCR sequences used by most existing models. Previous approaches depend on costly, inconsistent human annotations, limiting scalability and generalization. To bridge the gap, we propose a cost-effective paradigm that leverages large language models (LLMs) to infer global RO and inter-element layout relations without human supervision. By explicitly incorporating RO as structural guidance, our method captures hierarchical, document-level dependencies beyond local adjacency. Experiments on Semantic Entity Recognition, Entity Linking, and Document Question Answering show consistent improvements over baseline methods. Notably, LLM-inferred RO, even when differing from ground-truth adjacency, provides richer global structural priors and yields superior downstream performance. These results and findings demonstrate the scalability and significance of RO-aware modeling, advancing both LLMs and lightweight layout-aware models for robust document understanding. Code, data, and more details will be made publicly available after corporate review, in accordance with SAP’s corporate open-source policy.

Counterspeech generation has emerged as a promising approach to combat online hate speech, with recent work focusing on controlling attributes used in counterspeech, such as strategies or intents. While these attributes are often evaluated automatically using classifiers, a key goal of this evaluation is to compare the performance of different generation models. However, the validity of such evaluation results is questionable when the classifiers themselves have only modest performance. This paper examines the automatic evaluation of counterspeech attributes using a multi-attribute counterspeech dataset containing 2,728 samples. We investigate when automatic evaluation can be trusted for model comparison and address the limitations of current evaluation methodologies. We make concrete recommendations for how to perform classifier validation before model evaluation. Our classifier validation results demonstrate that even limited classifiers can produce trustworthy model rankings. Therefore, we argue that when comparing counterspeech generation models, a classifier’s ability to rank generation models is a more direct measure of its practical utility than traditional classification metrics, e.g., accuracy and F1.

Reasoning-augmented vision language models (VLMs) generate explicit chains of thought that promise greater capability and transparency but also introduce new failure modes: models may reach correct answers via visually unfaithful intermediate steps, or reason faithfully yet fail on the final prediction. Standard evaluations that only measure final-answer accuracy cannot distinguish these behaviors.We introduce the visual faithfulness of reasoning chains as a distinct evaluation dimension, focusing on whether the perception steps of a reasoning chain are grounded in the image. We propose a training- and reference-free framework that decomposes chains into perception versus reasoning steps and uses off-the-shelf VLM judges for step-level faithfulness, additionally verifying this approach through a human meta-evaluation. Building on this metric, we present a lightweight self-reflection procedure that detects and locally regenerates unfaithful perception steps without any training. Across multiple reasoning-trained VLMs and perception-heavy benchmarks, our method reduces Unfaithful Perception Rate while preserving final-answer accuracy, improving the reliability of multimodal reasoning.

Android is the largest mobile platform, yet automatically building applications remains a practical challenge. While Large Language Models (LLMs) show promise for code repair, their use for fixing Android build errors remains underexplored. To address this gap, we first introduce AndroidBuildBench, a benchmark of 1,019 build failures curated from the commit histories of 43 open-source Android projects. Each problem is paired with a verified solution from a subsequent commit, ensuring that fixes are feasible. Second, we propose GradleFixer, an LLM agent with domain-specific tools for inspecting and manipulating the Gradle build environment. GradleFixer achieves a resolve rate of 81.4% (pass@1), significantly outperforming a state-of-the-art coding agent that relies on a general-purpose shell. GradleFixer’s success suggests that while LLMs possess the high-level knowledge to solve these failures, they struggle to translate this knowledge into effective low-level actions using a general-purpose shell. We demonstrate the effectiveness of a strategy we term *Tool Bridging*, which replaces general-purpose shell commands with domain-aware abstractions. We hypothesize this approach works through two mechanisms: 1) it provides tools in an API-like format that LLMs use more reliably, and 2) it constrains the action space to relevant operations. This approach bridges the gap between the model’s high-level reasoning and effective low-level execution.

Leaderboards are crucial in the machine learning (ML) domain for benchmarking and tracking progress. However, creating leaderboards traditionally demands significant manual effort. In recent years, efforts have been made to automate leaderboard generation, but existing datasets for this purpose are limited by capturing only the best results from each paper and limited metadata. We present MetaLead, a fully human-annotated ML Leaderboard dataset that captures all experimental results for result transparency and contains extra metadata, such as the result experimental type: baseline, proposed method, or variation of proposed method for experiment-type guided comparisons, and explicitly separates train and test dataset for cross-domain assessment. This enriched structure makes MetaLead a powerful resource for more transparent and nuanced evaluations across ML research. MetaLead dataset and code repository: https://anonymous.4open.science/r/metalead-7CA3

Generative medical vision-language models (Med-VLMs) are primarily designed to generate complex textual information (e.g., diagnostic reports) from multimodal inputs including vision modality (e.g., medical images) and language modality (e.g., clinical queries). However, their security vulnerabilities remain underexplored. Med-VLMs should be capable of rejecting harmful queries, such as Provide detailed instructions for using this CT scan for insurance fraud. At the same time, addressing security concerns introduces the risk of over-defense, where safety-enhancing mechanisms may degrade general performance, causing Med-VLMs to reject benign clinical queries. In this paper, we propose a novel inference-time defense strategy to mitigate harmful queries, enabling defense against visual and textual jailbreak attacks. Using diverse medical imaging datasets collected from nine modalities, we demonstrate that our defense strategy based on synthetic clinical demonstrations enhances model safety without significantly compromising performance. Additionally, we find that increasing the demonstration budget alleviates the over-defense issue. We then introduce a mixed demonstration strategy as a trade-off solution for balancing security and performance under few-shot demonstration budget constraints. Warning: This paper contains content that may be deemed harmful.

Hateful speech detection is a key component of content moderation, yet current evaluation frameworks rarely assess why a text is deemed hateful. We introduce , a four-component metric suite designed to evaluate the reasoning quality of model explanations. It assesses (i) conclusion explicitness, (ii) faithfulness and causal grounding of quoted spans, (iii) protected group identification (policy-configurable), and (iv) logical consistency among these elements. Evaluated on six diverse hate speech datasets, reveals interpretability failures and annotation inconsistencies that are invisible to standard metrics like Accuracy or F1. Moreover, human evaluation shows strong agreement with , validating it as a practical tool for trustworthy and transparent moderation. Disclaimer: This paper contains sensitive content that may be disturbing to some readers.

We investigate how independent demographic bias mechanisms are from general demographic recognition in language models. Using a multi-task evaluation setup where demographics are associated with names, professions, and education levels, we measure whether models can be debiased while preserving demographic detection capabilities. We compare attribution-based and correlation-based methods for locating bias features. We find that targeted sparse autoencoder feature ablations in Gemma-2-9B reduce bias without degrading recognition performance: attribution-based ablations mitigate race and gender profession stereotypes while preserving name recognition accuracy, whereas correlation-based ablations are more effective for education bias. Qualitative analysis further reveals that removing attribution features in education tasks induces “prior collapse”, thus increasing overall bias. This highlights the need for dimension-specific interventions. Overall, our results show that demographic bias arises from task-specific mechanisms rather than absolute demographic markers, and that mechanistic inference-time interventions can enable surgical debiasing without compromising core model capabilities.

User simulation has long played a vital role in computer science due to its potential to support a wide range of applications. Language, as the primary medium of human communication, forms the foundation of social interaction and behavior. Consequently, simulating conversational behavior has become a key area of study. Recent advancements in large language models (LLMs) have significantly catalyzed progress in this domain by enabling high-fidelity generation of synthetic user conversation. In this paper, we survey recent advancements in LLM-based conversational user simulation. We introduce a novel taxonomy covering user granularity and simulation objectives. Additionally, we systematically analyze core techniques and evaluation methodologies. We aim to keep the research community informed of the latest advancements in conversational user simulation and to further facilitate future research by identifying open challenges and organizing existing work under a unified framework.

Offensive language detection poses a significant challenge in modern social spaces, necessitating advanced solutions. Online media platforms have been known to escalate acts of violence and broader conflicts, and thus, an automated system to help counter offensive content is essential. Traditional NLP models have typically dominated the field of hate speech detection, but require careful model design and extensive tuning. Moreover, a notable resource gap exists for addressing offensive languages, particularly those transcribed in non-native scripts, such as Roman Urdu and Urdu. This study explores the potential of pre-trained LLMs in using prompt-based methods using different transcriptions of the Urdu language, particularly their efficacy in detecting offensive content in diverse linguistic contexts. Our study employs state-of-the-art open-source LLMs, including advanced variants of Llama, Qwen, Lughaat, and proprietary GPT-4, which are evaluated through prompting strategies in different under-resourced languages. Our findings show that pre-trained LLMs achieve performance comparable to traditional fine-tuned benchmarks in detecting hateful and offensive content.

Multi-modal large language models (MLLMs) have recently shown impressive capabilities but are also highly vulnerable to jailbreak attacks. While white-box methods can generate adversarial visual inputs via gradient-based optimization, such approaches fail in realistic black-box settings where model parameters are inaccessible. Zeroth-order (ZO) optimization offers a natural path for black-box attacks by estimating gradients from queries, yet its application to MLLMs is challenging due to sequence-conditioned objectives, limited feedback, and massive model scales. To address these issues, we propose Zer0-Jack, the first direct black-box jailbreak framework for MLLMs based on ZO optimization. Zer0-Jack focuses on generating malicious images and introduces a patch-wise block coordinate descent strategy that stabilizes gradient estimation and reduces query complexity, enabling efficient optimization on billion-scale models. Experiments show that Zer0-Jack achieves 98.2% success on MiniGPT-4 and 95% on the Harmful Behaviors Multi-modal dataset, while directly jailbreaking commercial models such as GPT-4o. These results demonstrate that ZO optimization can be effectively adapted to jailbreak large-scale multi-modal LLMs. Codes are provided here.

Retrieving the biological impacts of protein-protein interactions (PPIs) is essential for target identification (Target ID) in drug development. Given the vast number of proteins involved, this process remains time-consuming and challenging. Large Language Models (LLMs) and Retrieval-Augmented Generation (RAG) frameworks have supported Target ID; however, no benchmark currently exists for identifying the biological impacts of PPIs. To bridge this gap, we introduce the RAG Benchmark for PPIs (RAGPPI), a factual question-answer benchmark of 4,420 question-answer pairs that focus on the potential biological impacts of PPIs. Through interviews with experts, we identified criteria for a benchmark dataset, such as a type of QA and source. We built a gold-standard dataset (500 QA pairs) through expert-driven data annotation. We developed an ensemble auto-evaluation LLM that incorporates expert labeling characteristics, average fact–abstract similarity (F₁), and low-similarity fact counts (F₂), enabling the construction of a silver-standard dataset (3,720 QA pairs). We are committed to maintaining RAGPPI as a resource to support the research community in advancing RAG systems for drug discovery QA solutions.

Large language models (LLMs) have demonstrated strong performance across diverse natural language processing tasks. However, their performance varies significantly across different prompts, requiring careful engineering for consistent results. Manual prompt engineering requires substantial human effort and suffers from limited reproducibility. In contrast, automatic prompt optimization methods reduce manual effort but often depend on costly autoregressive generation, resulting in substantial latency overheads. To address these limitations, we present low-latency prompt optimization (LLPO), a novel framework that reframes prompt engineering as a classification problem. LLPO classifies structured prompt fields from user input through multi-task classification and populates a predefined template to generate an optimized system prompt with minimal latency. In LLM-based automatic evaluations across four question-answering benchmarks, LLPO improves answer quality by up to 26.5% in ∆win rate compared to prior automatic prompt optimization methods, while reducing latency by up to 1,956 times. Human evaluation shows that LLPO receives the highest proportion of top-ranked responses. Furthermore, we analyze the contribution of each structured prompt field to performance, highlighting the robustness of our framework.

Large-scale vision–language models (LVLMs) have achieved remarkable progress on various reasoning tasks. However, most studies focus on natural photographic images and pay limited attention to multi-panel visual narratives such as comics. This leaves a clear gap in our understanding of how well LVLMs perform chronological reasoning across comic panels. To address this, we introduce **ChrOMIC**, a new benchmark dataset for **chro**nological reasoning in multi-panel **comic**s. It covers six types of reasoning questions and spans both Western and Japanese comic styles. To ensure high-quality annotations, we customized a human–AI collaborative annotation process tailored to the characteristics of the two comic styles. We further introduce three core tasks: Description Reordering and Panel Reordering, which jointly assess models’ ability to understand chronological order in panel sequences, and Multiple-Choice Question Answering (MCQA), which evaluates narrative-level reasoning. We evaluate a range of open-source and commercial LVLMs on ChrOMIC, and find that even the leading models struggle with panel-based chronological reasoning. Further analysis reveals key limitations, including weak visual action understanding and frequent hallucinations in fine-grained visual interpretation.

Parameter-Efficient Fine-Tuning (PEFT) has become a key strategy for adapting large language models, with recent advances in sparse tuning reducing overhead by selectively updating key parameters or subsets of data. Existing approaches generally focus on two distinct paradigms: layer-selective methods aiming to fine-tune critical layers to minimize computational load, and data-selective methods aiming to select effective training subsets to boost training. However, current methods typically overlook the fact that different data points contribute varying degrees to distinct model layers, and they often discard potentially valuable information from data perceived as of low quality. To address these limitations, we propose Gradient-aligned Sparse Tuning (GAST), an innovative method that simultaneously performs selective fine-tuning at both data and layer dimensions as integral components of a unified optimization strategy. GAST specifically targets redundancy in information by employing a layer-sparse strategy that adaptively selects the most impactful data points for each layer, providing a more comprehensive and sophisticated solution than approaches restricted to a single dimension. Experiments demonstrate that GAST consistently outperforms baseline methods, establishing a promising direction for future research in PEFT strategies.

We introduce BeDiscovER (Benchmark of Discourse Understanding in the Era of Reasoning Language Models), an up-to-date, comprehensive suite for evaluating the discourse-level knowledge of modern LLMs. BeDiscovER compiles 5 publicly available discourse tasks across discourse lexicon, (multi-)sentential, and documental levels, with in total 52 individual datasets. It covers both extensively studied tasks such as discourse parsing and temporal relation extraction, as well as some novel challenges such as discourse particle disambiguation (e.g., just), and also aggregates a shared-task on Discourse Relation Parsing and Treebanking for multilingual and multi-framework discourse relation classification. We evaluate open-source LLMs: Qwen3 series, DeepSeek-R1, and frontier reasoning model GPT-5-mini on BeDiscovER, and find that state-of-the-art models exhibits strong performance in arithmetic aspect of temporal reasoning, but they struggle with long-dependency reasoning and some subtle semantic and discourse phenomena, such as rhetorical relation classification.

Large language models (LLMs) demonstrate superior reasoning capabilities compared to small language models (SLMs), but incur substantially higher costs. We propose COllaborative REAsoner (COREA), a system that cascades an SLM with an LLM to achieve a balance between accuracy and cost in complex reasoning tasks. COREA first attempts to answer questions using the SLM, which outputs both an answer and a verbalized confidence score. Questions with confidence below a predefined threshold are deferred to the LLM for more accurate resolution. We introduce a reinforcement learning-based training algorithm that aligns the SLM’s confidence through an additional confidence calibration reward. Extensive experiments demonstrate that our method jointly improves the SLM’s reasoning ability and confidence calibration across diverse datasets and model backbones. Compared to using the LLM alone, COREA reduces cost by 21.5% and 16.8% on out-of-domain math and non-math datasets, respectively, with only an absolute pass@1 drop within 2%.

Ghosting, the ability to predict a user’s intended text input for inline query auto-completion, is an invaluable feature for modern search engines and chat interfaces, greatly enhancing user experience. By suggesting completions to incomplete queries (or prefixes), ghosting aids users with slow typing speeds, disabilities, or limited language proficiency. Ghosting is a challenging problem and has become more important with the ubiquitousness of chat-based systems like ChatGPT, Copilot, etc. Despite the increasing prominence of chat-based systems utilizing ghosting, this challenging problem of Chat-Ghosting has received little attention from the NLP/ML research community. There is a lack of standardized benchmarks and relative performance analysis of deep learning and non-deep learning methods. We address this through an open and thorough study of this problem using four publicly available dialog datasets: two human-human (DailyDialog and DSTC7-Ubuntu) and two human-bot (Open Assistant and ShareGPT). We experiment with various existing query auto-completion methods (using tries), n-gram methods and deep learning methods, with and without dialog context. We also propose a novel entropy-based dynamic early stopping strategy. Our analysis finds that statistical n-gram models and tries outperform deep learning based models in terms of both model performance and inference efficiency for seen prefixes. For unseen queries, neural models like T5 and Phi-2 lead to better results. Adding conversational context leads to significant improvements in ghosting quality, especially for Open-Assistant and ShareGPT. We make code and data publicly available at https://github.com/blitzprecision/Chat-Ghosting.

Vision-Language Models excel in multi-modal tasks but often hallucinate objects or exhibit linguistic bias by over-repeating object names, especially in complex multi-object scenes. Existing methods struggle with multi-object grounding because language priors frequently dominate visual evidence, causing hallucinated or biased objects to produce attention distributions or similarity scores nearly indistinguishable from those of real objects. We introduce SHAPLENS, a Shapley value–based attribution framework using Kernel SHAP and multi-layer fusion to detect hallucinated and biased objects. Evaluated on ADE and COCO datasets across four leading VLMs, SHAPLENS improves hallucination detection accuracy by 8–12% and F1 by 10–14% over the best baselines. It also achieves up to 6% higher bias detection performance across three distinct bias types on a curated HQH benchmark and exhibits minimal degradation (<0.03%) across partial and perturbed contexts.

Large language models (LLMs) continue to achieve impressive performance on reasoning benchmarks, yet it remains unclear how their predictions capture semantic consistency between sentences. We investigate the important open question of whether word-level surprisal correlates with sentence-level contradiction between a premise and a hypothesis. Specifically, we compute surprisal for hypothesis words across a diverse set of experimental variants, and analyze its association with contradiction labels over multiple datasets and open-source LLMs. Because modern LLMs operate on subword tokens and can not directly produce reliable surprisal estimates, we introduce a token-to-word decoding algorithm that extends theoretically grounded probability estimation to open-vocabulary settings. Experiments show a consistent and statistically significant positive correlation between surprisal and contradiction across models and domains. Our analysis also provides new insights into the capabilities and limitations of current LLMs. Together, our findings suggest that surprisal can localize sentence-level inconsistency at the word level, establishing a quantitative link between lexical uncertainty and sentential semantics. We plan to release our code and data upon publication.

Human cognition, driven by complex neurochemical processes, oscillates between imagination and reality and learns to self-correct whenever such subtle drifts lead to hallucinations or unsafe associations. In recent years, Large Language Models (LLMs) have demonstrated remarkable performance in a wide range of tasks. However, they still lack human cognition to balance factuality and safety. Bearing the resemblance, we argue that both factual and safety failures in LLMs arise from a common underlying issue, "representational misalignment" in their latent activation space. We hypothesize that an external network, trained to understand the fluctuations, can selectively intervene in the model to regulate falsehood into truthfulness and unsafe output into safe output without fine-tuning the LLM’s parameters. Reflecting the hypothesis, we propose ARREST (Adversarial Resilient Regulation Enhancing Safety and Truth), a unified framework that identifies and corrects drifted features, engaging both soft and hard refusals in addition to factual corrections. Our empirical results show that ARREST not only regulates misalignment but is also more versatile compared to the Reinforcement Learning from Human Feedback (RLHF)-aligned models in generating soft refusals due to adversarial training. We make our codebase available at https://github.com/sharanya-dasgupta001/ARREST.

Document-level Joint Entity and Relation Extraction (JERE) benchmarks such as DocRED, Re-DocRED, and DocGNRE suffer from pervasive False Negatives (FN), undermining training and evaluation. In this paper, we introduce SiftingLogic – a training-free annotation pipeline that leverages LLMs with user-specifiable reasoning, enriched inverse/co-occurring relation schemas, and novel entity-level constraints to systematically address FN gaps. Applying SiftingLogic and our enriched schema of inverse and co-occurring relations, we add 29,580 verified triplets to Re-DocRED (train/dev, +27%) and over 9,700 verified triplets to DocGNRE test (+49.89%), yielding the enhanced Re²-DocRED dataset. Beyond English datasets, we also apply our SiftingLogic to REDFM Mandarin test set, resulting in a significant increase in triplets from 663 to 1,391 (+109.8%) demonstrating our pipeline’s generalisability across languages and datasets. Our experiments show that recall scores of models trained on existing public datasets drop notably on our revised splits, whereas our enriched training set mitigates this, underscoring persistent FN gaps and motivating our proposed SiftingLogic and Re²-DocRED. To facilitate further research and reproducibility of our work, the Re²-DocRED dataset is released at https://github.com/klassessg/re2-docred.

Understanding how large language models (LLMs) process compositional linguistic structures is integral to enhancing their reliability and interpretability. We present Constituent-Aware Pooling (CAP), a methodology grounded in compositionality, mechanistic interpretability, and information theory that intervenes in model activations by pooling token representations into linguistic constituents at various layers. Experiments across eight models (124M-8B parameters) on inverse definition modelling, hypernym and synonym prediction reveal that semantic composition is not localised to specific layers but distributed across network depth. Performance degrades substantially under constituent-based pooling, particularly in early and middle layers, with larger models showing greater sensitivity. We propose an information-theoretic interpretation: transformers’ training objectives incentivise deferred integration to maximise token-level throughput, resulting in fragmented rather than localised composition. These findings highlight fundamental architectural and training constraints requiring specialised approaches to encourage robust compositional processing.

As vision-language models (VLMs) are deployed globally, their ability to understand culturally situated knowledge becomes essential. Yet, existing evaluations largely assess static recall or isolated visual grounding, leaving unanswered whether VLMs possess robust and transferable cultural understanding. We introduce ‘BLEnD-Vis‘, a multimodal, multicultural benchmark designed to evaluate the robustness of everyday cultural knowledge in VLMs across linguistic rephrasings and visual modalities. Building on the BLEnD dataset, ‘BLEnD-Vis‘ constructs 313 culturally grounded question templates spanning 16 regions and generates three aligned multiple-choice formats: (i) a text-only baseline querying from Region → Entity, (ii) an inverted text-only variant (Entity → Region), and (iii) a VQA-style version of (ii) with generated images. The resulting benchmark comprises 4,916 images and over 21,000 multiple-choice questions (MCQ) instances, validated through human annotation. ‘BLEnD-Vis‘ reveals significant fragility in current VLM cultural knowledge; models exhibit performance drops under linguistic rephrasing. While visual cues often aid performance, low cross-modal consistency highlights the challenges of robustly integrating textual and visual understanding, particularly in lower-resource regions. ‘BLEnD-Vis‘ thus provides a crucial testbed for systematically analysing cultural robustness and multimodal grounding, exposing limitations and guiding the development of more culturally competent VLMs. Code is available at https://github.com/Social-AI-Studio/BLEnD-Vis.

Document-Level Zero-Shot Relation Extraction (DocZSRE) aims to predict unseen relation labels in text documents without prior training on specific relations. Existing approaches rely on Large Language Models (LLMs) to generate synthetic data for unseen labels, which poses challenges for low-resource languages like Malaysian English. These challenges include the incorporation of local linguistic nuances and the risk of factual inaccuracies in LLM-generated data. This paper introduces Document-Level Zero-Shot Relation Extraction with Entity Side Information (DocZSRE-SI) to address limitations in the existing DocZSRE approach. The DocZSRE-SI framework leverages Entity Side Information, such as Entity Mention Descriptions and Entity Mention Hypernyms, to perform ZSRE without depending on LLM-generated synthetic data. The proposed low-complexity model achieves an average improvement of 11.6% in the macro F1-Score compared to baseline models and existing benchmarks. By utilising Entity Side Information, DocZSRE-SI offers a robust and efficient alternative to error-prone, LLM-based methods, demonstrating significant advancements in handling low-resource languages and linguistic diversity in relation extraction tasks. This research provides a scalable and reliable solution for ZSRE, particularly in contexts like Malaysian English news articles, where traditional LLM-based approaches fall short.

Large language models have simplified the production of personalized translations reflecting predefined stylistic constraints. However, these systems still struggle when stylistic requirements are implicitly represented by a set of examples, such as texts produced by a specific human translator. In this work, we explore various strategies for personalizing automatically generated translations when few examples are available, with a focus on the challenging domain of literary translation. We begin by determining the feasibility of the task and how style information is encoded within model representations. Then, we evaluate various prompting strategies and inference-time interventions for steering model generations towards a personalized style, with a particular focus on contrastive steering with sparse autoencoder (SAE) latents to identify salient personalization properties. We demonstrate that contrastive SAE steering yields robust style conditioning and translation quality, resulting in higher inference-time computational efficiency than prompting approaches. We further examine the impact of steering on model activations, finding that layers encoding personalization properties are impacted similarly by prompting and SAE steering, suggesting a similar mechanism at play.

How capable are large language models (LLMs) in the domain of taxation? Although numerous studies have explored the legal domain, research dedicated to taxation remains scarce. Moreover, the datasets used in these studies are either simplified, failing to reflect the real-world complexities, or not released as open-source. To address this gap, we introduce PLAT, a new benchmark designed to assess the ability of LLMs to predict the legitimacy of additional tax penalties. PLAT comprises 300 examples: (1) 100 binary-choice questions, (2) 100 multiple-choice questions, and (3) 100 essay-type questions, all derived from 100 Korean court precedents. PLAT is constructed to evaluate not only LLMs’ understanding of tax law but also their performance in legal cases that require complex reasoning beyond straight forward application of statutes. Our systematic experiments with multiple LLMs reveal that (1) their baseline capabilities are limited, especially in cases involving conflicting issues that require a comprehensive understanding (not only of the statutes but also of the taxpayer’s circumstances), and (2) LLMs struggle particularly with the “AC” stages of “IRAC” even for advanced reasoning models like o3, which actively employ inference-time scaling.

Knowledge editing (KE) has recently emerged as a promising technique to update specific facts in large language models (LLMs) without full retraining. While existing KE methods show promising results on general-domain benchmarks, their effectiveness in the medical domain remains largely unexplored. Medical knowledge editing poses unique challenges, requiring models not only to memorize new facts but also to internalize and generalize them for reliable and interpretable clinical decision-making. In this work, we propose MedEditBench, a rigorous evaluation framework for assessing medical knowledge editing. Our preliminary results reveal that current KE paradigm, which directly edits simple answers to the LLMs, often leads to superficial updates with poor generalization. To address this, we introduce Self-Generated Rationale Editing (SGR-Edit), which leverages model-generated rationales as editing targets, enabling deeper knowledge integration. Extensive experiments across diverse LLMs and KE methods demonstrate that SGR-Edit consistently improves editing efficacy and generalization. Furthermore, we examine the impact of sequential edits on in-domain medical knowledge, external-domain knowledge, as well as general model capabilities, offering practical insights for deploying KE in real-world medical applications.

Large language models (LLMs) have emerged as strong contenders in machine translation. Yet, they still struggle to adequately handle discourse phenomena, such as pronoun resolution and lexical cohesion at the document level. In this study, we thoroughly investigate the discourse phenomena performance of LLMs in context-aware translation. We demonstrate that discourse knowledge is encoded within LLMs and propose the use of quality-aware decoding (QAD), specifically minimum Bayes risk decoding, to effectively extract this knowledge, showcasing its superiority over other decoding approaches through comprehensive analysis. Furthermore, we illustrate that QAD enhances the semantic richness of translations and aligns them more closely with human preferences.

Grammaticalization denotes a diachronic change of the grammatical category from content words to function words. One of the intensively explored directions in this area is to quantify the degree of grammaticalization. There have been a limited number of automated methods for this task and the existing, best-performing method is heavily language- and word-dependent. In this paper, we explore three methods for quantifying the degree of grammaticalization, which are applicable to a wider variety of words and languages. The difficulty here is that training data is not available in the present task. We overcome this difficulty by using Positive-Unlabeled learning (PU-learning) or Cross-Validation-like learning (hereafter, CV-learning). Experiments show that the CV-learning-based method achieves middle to high correlations to human judgments in English deverbal prepositions and Japanese nouns being grammaticalized. With this method, we further explore words possibly being grammaticalized and counterexamples of the unidirectionality hypothesis.

Large Language Models (LLMs) are increasingly used for knowledge-based reasoning tasks, yet understanding when they rely on genuine knowledge versus superficial heuristics remains challenging. We investigate this question through entity comparison tasks by asking models to compare entities along numerical attributes (e.g., “Which river is longer, the Danube or the Nile?”), which offer clear ground truth for systematic analysis. Despite having sufficient numerical knowledge to answer correctly, LLMs frequently make predictions which contradict this knowledge. We identify three heuristic biases that strongly influence model predictions: entity popularity, mention order, and semantic co-occurrence. For smaller models, a simple logistic regression using only these surface cues predicts model choices more accurately than the model’s own numerical predictions, suggesting heuristics largely override principled reasoning. Crucially, we find that larger models (32B parameters) selectively rely on numerical knowledge when it is more reliable, while smaller models (7-8B parameters) show no such discrimination, which explains why larger models outperform smaller ones even when the smaller models possess more accurate knowledge. Chain-of-thought prompting steers all models towards using the numerical features across all model sizes.

Quantization is an effective technique for reducing the storage footprint and computational costs of Large Language Models (LLMs), but it often results in performance degradation. Existing post-training quantization methods typically use small, English-only calibration sets; however, their impact on multilingual models remains underexplored. We systematically evaluate eight calibration settings (five single-language and three multilingual mixes) across two quantizers (GPTQ, AWQ) on data from 10 different languages. Our findings reveal a consistent trend: non-English and multilingual calibration sets significantly improve perplexity compared to English-only baselines. Specifically, we observe notable average perplexity gains across both quantizers on Llama3.1 8B and Qwen2.5 7B, with multilingual mixes achieving the largest overall reductions of up to 3.52 perplexity gain. Furthermore, our analysis indicates that tailoring calibration sets to the evaluation language yields the largest improvements for individual languages, underscoring the importance of linguistic alignment. We also identify specific failure cases where certain language-quantizer combinations degrade performance, which we trace to differences in activation range distributions across languages. These results highlight that static, one-size-fits-all calibration is suboptimal, and that tailoring calibration data, both in language and diversity, plays a crucial role in robustly quantizing multilingual LLMs.

Large Language Models (LLMs) have achieved impressive performance yet remain inconsistent across languages, often defaulting to high-resource outputs such as English. Existing multilingual alignment methods mitigate these issues through preference optimization but rely on external supervision, such as translation systems or English-biased signal. We propose Multilingual Self-Alignment (MSA), a targeted preference optimization framework that leverages an LLM’s own latent representations as intrinsic supervision signals, rewarding lower-resource language outputs based on their alignment with high-resource (English) counterparts in the "semantic hub". We further introduce Language-Consistency MSA (LaCoMSA), which augments MSA with a final-layer language-consistency factor to prevent off-target generation. Integrated with Direct Preference Optimization, LaCoMSA improves a Llama 3 8B-based model multilingual win rates by up to 6.8% absolute (55.0% relatively) on X-AlpacaEval and achieves consistent gains across benchmarks and models. Our findings demonstrate that LaCoMSA can serve as an effective and scalable mechanism, opening a new venue toward multilingual self-alignment.

Large language models (LLMs) can answer prompts in many languages, despite being trained predominantly on English; yet, the mechanisms driving this generalization remain poorly understood. This work asks: How does an LLM’s ability to align representations of non-English inputs to English impact its performance on natural language understanding (NLU) tasks? We study the role of representation alignment in instance-level task decisions, complementing prior analyses conducted both at the language level and task-independently. We introduce the Discriminative Alignment Index (\DALI) to quantify instance-level alignment across 24 languages other than English and three distinct NLU tasks. Results show that incorrect NLU predictions are strongly associated with lower representation alignment with English in the model’s middle layers. Through activation patching, we show that incorrect predictions in languages other than English can be fixed by patching their parallel English activations in the middle layers, thereby demonstrating the causal role of representation (mis)alignment in cross-lingual correctness.

Much recent work has shown how cross-linguistic variation is constrained by competing pressures from efficient communication. However, little attention has been paid to the role of the systematicity of forms (*regularity*), a key property of natural language. Here, we demonstrate the importance of regularity in explaining the shape of linguistic systems by looking at recursive numeral systems. Previous work has argued that these systems optimise the trade-off between lexicon size and average morphosyntatic complexity (Denić and Szymanik, 2024). However, showing that *only* natural-language-like systems optimise this trade-off has proven elusive, and existing solutions rely on ad-hoc constraints to rule out unnatural systems (Yang and Regier, 2025). Drawing on the Minimum Description Length (MDL) approach, we argue that recursive numeral systems are better viewed as efficient with regard to their regularity and processing complexity. We show that our MDL-based measures of regularity and processing complexity better capture the key differences between attested, natural systems and theoretically possible ones, including “optimal” recursive numeral systems from previous work, and that the ad-hoc constraintsnaturally follow from regularity. Our approach highlights the need to incorporate regularity across sets of forms in studies attempting tomeasure efficiency in language.

Pretrained Transformer encoders are the dominant approach to sequence labeling. While some alternative architectures-such as xLSTMs, structured state-space models, diffusion models, and adversarial learning-have shown promise in language modeling, few have been applied to sequence labeling, and mostly on flat or simplified tasks. We study how these architectures adapt across tagging tasks that vary in structural complexity, label space, and token dependencies, with evaluation spanning multiple languages. We find that the strong performance previously observed in simpler settings does not always generalize well across languages or datasets, nor does it extend to more complex structured tasks.

Large Language Models (LLMs) exhibit remarkable capabilities, but no single model optimally balances serving quality and deployment cost across diverse tasks. Multi-LLM systems address this challenge through intelligent routing mechanisms that dynamically allocate queries to the most appropriate model. However, existing routing methods suffer from two fundamental limitations: (i) dependence on extensive full-response datasets for training, and (ii) poor scalability when incorporating new models, typically necessitating retraining from scratch. In this paper, we propose SemiRouter, a novel LLM routing framework designed for data-sparse and evolving model environments. Our approach combines a data-efficient training methodology with an adaptive architecture that enables seamless integration of new models under limited supervision. As an extension, we also consider energy footprint as a potential deployment cost in our routing decision. Empirical evaluations demonstrate that our method improves data efficiency, adaptability, and routing accuracy compared to existing approaches, providing a scalable solution for dynamic multi-LLM deployment.

The promise of LLM watermarking rests on a core assumption that a specific watermark proves authorship by a specific model. We demonstrate that this assumption is dangerously flawed. We introduce the threat of watermark spoofing, a sophisticated attack that allows a malicious model to generate text containing the authentic-looking watermark of a trusted, victim model. This enables the seamless misattribution of harmful content, such as disinformation, to reputable sources. The key to our attack is repurposing watermark radioactivity, the unintended inheritance of data patterns during fine-tuning, from a discoverable trait into an attack vector. By distilling knowledge from a watermarked teacher model, our framework allows an attacker to steal and replicate the watermarking signal of the victim model. This work reveals a critical security gap in text authorship verification and calls for a paradigm shift towards technologies capable of distinguishing authentic watermarks from expertly imitated ones. Our code is available at https://github.com/hsannn/ditto.

Event argument extraction (EAE) is a crucial task in information extraction. However, its performance heavily depends on expensive annotated data, making data scarcity a persistent challenge. Data augmentation serves as an effective approach to improving model performance in low-resource settings, yet research on applying LLMs for EAE augmentation remains preliminary. In this study, we pay attention to the boundary sensitivity of EAE and investigate four LLM-based augmentation strategies: argument replacement, adjunction rewriting, their combination, and annotation generation. We conduct comprehensive experiments across four benchmark datasets, employing GPT-4o-Mini and DeepSeek-R1-7B as data generators. Our results show that boundary-aware augmentation consistently leads to greater performance improvements over boundary-agnostic methods. In addition to performance gains, we provide a detailed analysis of augmentation quality from multiple perspectives, including uncertainty reduction, error types, data quality, and data scale. This work offers both empirical evidence and practical guidance for leveraging LLMs to enhance event argument extraction under low-resource conditions.

The meaning conveyed by a sentence often depends on the context in which it appears. Despite the progress of sentence embedding methods, it remains unclear as how to best modify a sentence embedding conditioned on its context. To address this problem, we propose Condition-Aware Sentence Embeddings (CASE), an efficient and accurate method to create an embedding for a sentence under a given condition. First, CASE creates an embedding for the condition using an Large Language Model (LLM) encoder, where the sentence influences the attention scores computed for the tokens in the condition during pooling. Next, a supervised method is learnt to align the LLM-based text embeddings with the Conditional Semantic Textual Similarity (C-STS) task. We find that subtracting the condition embedding will consistently improve the C-STS performance of LLM-based text embeddings and improve the isotropy of the embedding space. Moreover, our supervised projection method significantly improves the performance of LLM-based embeddings despite requiring a small number of embedding dimensions.

ICD coding is the process of mapping unstructured text from Electronic Health Records (EHRs) to standardised codes defined by the International Classification of Diseases (ICD) system. In order to promote trust and transparency, existing explorations on the explainability of ICD coding models primarily rely on attention-based rationales and qualitative assessments conducted by physicians, yet lack a systematic evaluation across diverse types of rationales using consistent criteria and high-quality rationale-annotated datasets specifically designed for the ICD coding task. Moreover, dedicated methods explicitly trained to generate plausible rationales remain scarce. In this work, we present evaluations of the explainability of rationales in ICD coding, focusing on two fundamental dimensions: faithfulness and plausibility—in short how rationales influence model decisions and how convincing humans find them. For plausibility, we construct a novel, multi-granular rationale-annotated ICD coding dataset, based on the MIMIC-IV database and the updated ICD-10 coding system. We conduct a comprehensive evaluation across three types of ICD coding rationales: entity-level mentions automatically constructed via entity linking, LLM-generated rationales, and rationales based on attention scores of ICD coding models. Building upon the strong plausibility exhibited by LLM-generated rationales, we further leverage them as distant supervision signals to develop rationale learning methods. Additionally, by prompting the LLM with few-shot human-annotated examples from our dataset, we achieve notable improvements in the plausibility of rationale generation in both the teacher LLM and the student rationale learning models.

Retrieval Augmented Generation (RAG) is a popular method for injecting up-to-date information into Large Language Model (LLM)-based Natural Language Generation (NLG) systems. While RAG can enhance factual accuracy, its effect on the social biases inherent in LLMs is not well understood. This paper systematically investigates how RAG modulates social biases across three languages (English, Japanese, and Chinese) and four categories (gender, race, age, and religion). By evaluating various generator LLMs on the BBQ benchmark, we analyse how document collections with controlled stereotypical content affect RAG outputs. We find that biases present in the retrieved documents are often significantly amplified in the generated texts, even when the base LLM itself has a low-level of intrinsic bias. These findings raise concerns about the social fairness of RAG systems, underscoring the urgent need for careful bias evaluation before real-world deployment.

Existing challenges in misinformation exposure and susceptibility vary across demographics, as some populations are more vulnerable to misinformation than others. Large language models (LLMs) introduce new dimensions to these challenges through their ability to generate persuasive content at scale and reinforcing existing biases. Our study introduces PANDORA, a framework that investigates the bidirectional persuasion dynamics between LLMs and humans when exposed to misinformative content. We use a multi-agent LLM framework to analyze the spread of misinformation under persuasion among demographic-oriented LLM agents. Our findings show that demographic factors influence LLM susceptibility, with up to 15 percentage point differences in misinformation correctness across groups. Multi-agent LLMs also exhibit echo chamber behavior, aligning with human-like group polarization patterns. Therefore, this work highlights demographic divides in misinformation dynamics and offers insights for future interventions.

Test-time compute methods can significantly improve the reasoning capabilities and problem-solving accuracy of large language models. However, these approaches require substantially more computational resources, with most computation wasted on exploring low-diversity branches where the model already exhibits high confidence. We observe that a small subset of uncertain reasoning steps has a disproportionately large impact on final prediction accuracy, and branching at these points tends to yield higher-quality and more diverse candidate reasoning steps. Therefore, we introduce Entropy-Gated Branching: a novel inference technique that dynamically allocates computational resources by selectively expanding prediction sequences only at points of high uncertainty. Our method leverages entropy as a gating mechanism to identify when branching is most beneficial, coupled with an external feedback model to rank and prune candidate branches. Empirical results on mathematical and financial reasoning benchmarks show that this strategy improves accuracy by 22.6% over standard inference while operating 31%-75% faster across math benchmarks than test-time beam search with higher performance. Our results show that dynamic resource allocation during inference can substantially improve both efficiency and effectiveness, offering a more scalable pathway to enhanced LLM reasoning capabilities. We release our code and tools here[<https://github.com/JXL884/entropy_gated_branching>]

Large language models (LLMs) are increasingly used for long-document question answering, where reliable attribution to sources is critical for trust. Existing post-hoc attribution methods work well for extractive QA but struggle in multi-hop, abstractive, and semi-extractive settings, where answers synthesize information across passages. To address these challenges, we argue that post-hoc attribution can be reframed as a reasoning problem, where answers are decomposed into constituent units, each tied to specific context. We first show that prompting models to generate such decompositions alongside attributions improves performance. Building on this, we introduce DecompTune, a post-training method that teaches models to produce answer decompositions as intermediate reasoning steps. We curate a diverse dataset of complex QA tasks, annotated with decompositions by a strong LLM, and post-train Qwen-2.5 (7B and 14B) using a two-stage SFT + GRPO pipeline with task-specific curated rewards. Across extensive experiments and ablations, DecompTune substantially improves attribution quality, outperforming prior methods and matching or exceeding state-of-the-art frontier models.

Administrative phone tasks drain roughly $1 trillion annually from U.S. healthcare, with over 500 million insurance–benefit verification calls manually handled in 2024. We introduce INSURE-Dial, to our knowledge the first public benchmark for developing and assessing compliance-aware voice agents for phase-aware call auditing with span-based compliance verification. The corpus includes 50 de-identified, AI-initiated calls with live insurance representatives (mean 71 turns/call) and 1,000 synthetically generated calls that mirror the same workflow. All calls are annotated with a phase-structured JSON schema covering IVR navigation, patient identification, coverage status, medication checks (up to two drugs), and agent identification (CRN), and each phase is labeled for Information and Procedural compliance under explicit ask/answer logic. We define two novel evaluation tasks: (1) Phase Boundary Detection (span segmentation under phase-specific acceptance rules) and (2) Compliance Verification (IC/PC decisions given fixed spans). Per-phase scores are strong across small, low-latency baselines, but end-to-end reliability is constrained by span-boundary errors. On real calls, full-call exact segmentation is low, showing a gap between conversational fluency and audit-grade evidence.

Natural language processing (NLP) now shapes many aspects of our world, yet its potential for positive social impact is underexplored. This paper surveys work in “NLP for Social Good" (NLP4SG) across nine domains relevant to global development and risk agendas, summarizing principal tasks and challenges. We analyze ACL Anthology trends, finding that inclusion and AI harms attract the most research, while domains such as poverty, peacebuilding, and environmental protection remain underexplored. Guided by our review, we outline opportunities for responsible and equitable NLP and conclude with a call for cross-disciplinary partnerships and human-centered approaches to ensure that future NLP technologies advance the public good.

Large language models (LLMs) have shown promising accuracy in predicting survey responses and policy preferences, which has increased interest in their potential to represent human interests in various domains. Most existing research has focused on “behavioral cloning”, effectively evaluating how well models reproduce individuals’ expressed preferences. Drawing on theories of political representation, we highlight an underexplored design trade-off: whether AI systems should act as delegates, mirroring expressed preferences, or as trustees, exercising judgment about what best serves an individual’s interests. This trade-off is closely related to issues of LLM sycophancy, where models can encourage behavior or validate beliefs that may be aligned with a user’s short-term preferences, but is detrimental to their long-term interests. Through a series of experiments simulating votes on various policy issues in the U.S. context, we apply a temporal utility framework that weighs short and long-term interests (simulating a trustee role) and compare voting outcomes to behavior-cloning models (simulating a delegate). We find that trustee-style predictions weighted toward long-term interests produce policy decisions that align more closely with expert consensus on well-understood issues, but also show greater bias toward models’ default stances on topics lacking clear agreement. These findings reveal a fundamental trade-off in designing AI systems to represent human interests. Delegate models better preserve user autonomy but may diverge from well-supported policy positions, while trustee models can promote welfare on well-understood issues yet risk paternalism and bias.

Multilingual Large Language Models (LLMs) offer powerful capabilities for cross-lingual fact-checking. However, these models often exhibit language bias, performing disproportionately better on high-resource languages such as English than on low-resource counterparts. We also present and inspect a novel concept - retrieval bias, when information retrieval systems tend to favor certain information over others, leaving the retrieval process skewed. In this paper, we study language and retrieval bias in the context of Previously Fact-Checked Claim Detection (PFCD). We evaluate six open-source multilingual LLMs across 20 languages using a fully multilingual prompting strategy, leveraging the AMC-16K dataset. By translating task prompts into each language, we uncover disparities in monolingual and cross-lingual performance and identify key trends based on model family, size, and prompting strategy. Our findings highlight persistent bias in LLM behavior and offer recommendations for improving equity in multilingual fact-checking. To investigate retrieval bias, we employed multilingual embedding models and look into the frequency of retrieved claims. Our analysis reveals that certain claims are retrieved disproportionately across different posts, leading to inflated retrieval performance for popular claims while under-representing less common ones.

Figurative language, especially fixed figurative expressions (FFEs) such as idioms and proverbs, poses unique challenges for large language models (LLMs). Unlike literal phrases, FFEs are culturally grounded and often non-compositional, making them vulnerable to figurative hallucination, the generation or acceptance of plausible-sounding but culturally invalid expressions. We introduce FFE-Hallu, the first comprehensive benchmark for evaluating LLMs’ ability to generate, detect, and translate FFEs in Persian, a linguistically rich but underrepresented language. FFE-Hallu includes 600 carefully curated examples spanning three tasks: FFE generation from meaning, detection of fabricated FFEs (across four controlled categories), and FFE-to-FFE translation from English to Persian. Our evaluation of six state-of-the-art multilingual LLMs reveals persistent weaknesses in both cultural grounding and figurative competence. While models like GPT-4.1 display relative strength in rejecting fabricated FFEs and retrieving authentic ones, most systems struggle to reliably distinguish real FFEs from high-quality fabrications and often hallucinate in translation. This work shows that LLMs still have important gaps in understanding and using figurative language, and that specialized benchmarks like FFE-Hallu are needed.

Verifying the truthfulness of claims usually requires joint multi-modal reasoning over both textual and visual evidence, such as analyzing both textual caption and chart image for claim verification. In addition, to make the reasoning process transparent, a textual explanation is necessary to justify the verification result. However, most claim verification works mainly focus on the reasoning over textual evidence only or ignore the explainability, resulting in inaccurate and unconvincing verification. To address this problem, we propose a novel model that jointly achieves evidence retrieval, multi-modal claim verification, and explanation generation. For evidence retrieval, we construct a two-layer multi-modal graph for claims and evidence, where we design image-to-text and text-to-image reasoning for multi-modal retrieval. For claim verification, we propose token- and evidence-level fusion to integrate claim and evidence embeddings for multi-modal verification. For explanation generation, we introduce multi-modal Fusion-in-Decoder for explainability. Finally, since almost all the datasets are in general domain, we create a scientific dataset, AIChartClaim, in AI domain to complement claim verification community. Experiments show the strength of our model.

Arabic calligraphy represents one of the richest visual traditions of the Arabic language, blending linguistic meaning with artistic form. Although multimodal models have advanced across languages, their ability to process Arabic script, especially in artistic and stylized calligraphic forms, remains largely unexplored. To address this gap, we present DuwatBench, a benchmark of 1,272 curated samples containing about 1,475 unique words across 6 classical and modern calligraphic styles, each paired with sentence-level detection annotations. The dataset reflects real-world challenges in Arabic writing, such as complex stroke patterns, dense ligatures, and stylistic variations that often challenge standard text recognition systems.Using DuwatBench, we evaluated 13 leading Arabic and multilingual multimodal models and showed that while they perform well in clean text, they struggle with calligraphic variation, artistic distortions, and precise visual–text alignment. By publicly releasing DuwatBench and its annotations, we aim to advance culturally grounded multimodal research, foster fair inclusion of Arabic language and visual heritage in AI systems, and support continued progress in this area. Our dataset and code are publicly available.

The promise of *LLM-based user simulators* to improve conversational AI is hindered by a critical "realism gap," leading to systems that are optimized for simulated interactions, but may fail to perform well in the real world. We introduce *ConvApparel*, a new dataset of human-AI conversations designed to address this gap. Its unique dual-agent data collection protocol, using both "good" and "bad" recommenders, enables counterfactual validation by capturing a wide spectrum of user experiences, enriched with first-person annotations of user satisfaction.We propose a comprehensive validation framework that combines *statistical alignment*, a *human-likeness score*, and *counterfactual validation* to test for generalization.Our experiments reveal a significant realism gap across all simulators. However, the framework also shows that data-driven simulators outperform a prompted baseline, particularly in counterfactual validation where they adapt more realistically to unseen behaviors, suggesting they embody more robust, if imperfect, user models.

This paper presents a novel task of extracting low-resourced and noisy Latin fragments from mixed-language historical documents with varied layouts. We benchmark and evaluate the performance of large foundation models against a multimodal dataset of 724 annotated pages. The results demonstrate that reliable Latin detection with contemporary zero-shot models is achievable, yet these models lack a functional comprehension of Latin. This study establishes a comprehensive baseline for processing Latin within mixed-language corpora, supporting quantitative analysis in intellectual history and historical linguistics. Both the dataset and code are available at https://github.com/COMHIS/EACL26-detect-latin.

Persona-assigned large language models (LLMs) are used in domains such as education, healthcare, and sociodemographic simulation. Yet, they are typically evaluated only in short, single-round settings that do not reflect real-world usage. We introduce an evaluation protocol that combines long persona dialogues (over 100 rounds) and evaluation datasets to create dialogue-conditioned benchmarks that can robustly measure long-context effects. We then investigate the effects of dialogue length on persona fidelity, instruction-following, and safety of seven state-of-the-art open- and closed-weight LLMs. We find that persona fidelity degrades over the course of dialogues, especially in goal-oriented conversations, where models must sustain both persona fidelity and instruction following. We identify a trade-off between fidelity and instruction following, with non-persona baselines initially outperforming persona-assigned models; as dialogues progress and fidelity fades, persona responses become increasingly similar to baseline responses. Our findings highlight the fragility of persona applications in extended interactions and our work provides a protocol to systematically measure such failures.

With their growing capabilities, generative large language models (LLMs) are being increasingly investigated for complex medical tasks.However, their effectiveness in real-world clinical applications remains underexplored. To address this, we present CliniBench, the first benchmark that enables comparability of well-studied encoder-based classifiers and generative LLMs for discharge diagnosis prediction from admission notes in the MIMIC-IV dataset. Our extensive study compares 12 generative LLMs and 3 encoder-based classifiers and demonstrates that encoder-based classifiers consistently outperform generative models in diagnosis prediction. We assess several retrieval augmentation strategies for in-context learning from similar patients and find that they provide notable performance improvements for generative LLMs.

In this study, we present a multimodal framework for predicting neuro-facial disorders by capturing both vocal and facial cues. We hypothesize that explicitly disentangling shared and modality-specific representations within multimodal foundation model embeddings can enhance clinical interpretability and generalization. To validate this hypothesis, we propose DIVINE a fully disentangled multimodal framework that operates on representations extracted from state-of-the-art (SOTA) audio and video foundation models, incorporating hierarchical variational bottlenecks, sparse gated fusion, and learnable symptom tokens. DIVINE operates in a multitask learning setup to jointly predict diagnostic categories (Healthy Control,ALS, Stroke) and severity levels (Mild, Moderate, Severe). The model is trained using synchronized audio and video inputs and evaluated on the Toronto NeuroFace dataset under full (audio-video) as well as single-modality (audio-only and video-only) test conditions. Our proposed approach, DIVINE achieves SOTA result, with the DeepSeek-VL2 and TRILLssoncombination reaching 98.26% accuracy and 97.51% F1-score. Under modality-constrained scenarios, the framework performs well, show-ing strong generalization when tested with video-only or audio-only inputs. It consistently yields superior performance compared to uni-modal models and baseline fusion techniques. To the best of our knowledge, DIVINE is the first fully disentangled multimodal frameworkto jointly perform categorical diagnosis and severity estimation for oro-facial neurological disorders using synchronized speech and facialvideo.

Are large language models (LLMs) sensitive to the distinction between humanly possible and impossible languages? This question was recently used in a broader debate on whether LLMs and humans share the same innate learning biases. Previous work has answered it in the positive by comparing LLM learning curves on existing language datasets and on "impossible" datasets derived from them via various perturbation functions. Using the same methodology, we examine this claim on a wider set of languages and impossible perturbations. We find that in most cases, GPT-2 learns each language and its impossible counterpart equally easily, in contrast to previous findings. We also apply a more lenient condition by testing whether GPT-2 provides any kind of separation between the whole set of natural languages and the whole set of impossible languages, based on cross-linguistic variance in metrics derived from the learning curves. Taken together, these perspectives show that GPT-2 provides no systematic separation between the possible and the impossible.

We propose a unified framework for not only attributing synthetic speech to its source but also for detecting speech generated by synthesizers that were not encountered during training. This requires methods that move beyond simple detection to support both detailed forensic analysis and open-set generalization. To address this, we introduce SIGNAL, a hybrid framework that combines speech foundation models (SFMs) with graph-based modeling and open-set-aware inference. Our framework integrates Graph Neural Networks (GNNs) and a k-Nearest Neighbor (KNN) classifier, allowing it to capture meaningful relationships between utterances and recognize speech that doesn’t belong to any known generator. It constructs a query-conditioned graph over generator class prototypes, enabling the GNN to reason over relationships among candidate generators, while the KNN branch supports open-set detection via confidence-based thresholding. We evaluate SIGNAL using the DiffSSD dataset, which offers a diverse mix of real speech and synthetic audio from both open-source and commercial diffusion-based TTS systems. To further assess generalization, we also test on the SingFake benchmark. Our results show that SIGNAL consistently improves performance across both tasks, with Mamba-based embeddings delivering especially strong results. To the best of our knowledge, this is the first study to unify graph-based learning and open-set detection for tracing synthetic speech back to its origin.

As large language models (LLMs) are trained on increasingly vast and opaque text corpora, determining which data contributed to training has become essential for copyright enforcement, compliance auditing, and user trust. While prior work focuses on detecting whether a dataset was used in training (membership inference), the complementary problem—verifying that a dataset was not used- has received little attention. We address this gap by introducing PRISM, a test that detects dataset-level non-membership using only grey-box access to model logits. Our key insight is that two models that have not seen a dataset exhibit higher rank correlation in their normalized token log probabilities than when one model has been trained on that data. Using this observation, we construct a correlation-based test that detects non-membership. Empirically, PRISM reliably rules out membership in training data across all datasets tested while avoiding false positives, thus offering a framework for verifying that specific datasets were excluded from LLM training.

Multimodal Large Language Models (MLLMs) often suffer from hallucinations, particularly errors in object existence, attributes, or relations, which undermine their reliability. We introduce TACO (Verified Atomic Confidence Estimation), a simple framework that mitigates hallucinations through self-verification and confidence calibration without relying on external vision experts. TACO decomposes responses into atomic queries, paraphrases them to reduce sensitivity to wording, and estimates confidence using self-consistency (black-box) or self-confidence (gray-box) aggregation, before refining answers with a language model. Experiments on five benchmarks (POPE, MME, HallusionBench, AMBER, and MM-Hal Bench) with two MLLMs (LLaVA-1.5-7B and CogVLM2) show that TACO consistently outperforms direct prompting and Visual Contrastive Decoding, reduces systematic biases, and improves confidence calibration, demonstrating its effectiveness in enhancing the faithfulness of MLLMs.

Specialized visual tools can augment large language models or vision language models with expert knowledge (e.g., grounding, spatial reasoning, medical knowledge, etc.), but knowing which tools to call (and when to call them) can be challenging. We introduce DART, a multi-agent framework that uses disagreements between multiple debating visual agents to identify useful visual tools (e.g., object detection, OCR, spatial reasoning, etc.) that can resolve inter-agent disagreement. These tools allow for fruitful multi-agent discussion by introducing new information, and by providing tool-aligned agreement scores that highlight agents in agreement with expert tools, thereby facilitating discussion. We utilize an aggregator agent to select the best answer by providing the agent outputs and tool information. We test DART on four diverse benchmarks and show that our approach improves over multi-agent debate as well as over single agent tool-calling frameworks, beating the next-strongest baseline (multi-agent debate with a judge model) by 3.4% and 2.4% on A-OKVQA and MMMU respectively. We also find that DART adapts well to new tools in applied domains, with a 1.3% improvement on the M3D medical dataset over other strong tool-calling, single agent, and multi-agent baselines. Additionally, we measure text overlap across rounds to highlight the rich discussion in DART compared to existing multi-agent methods. Finally, we study the distribution of expert tool calls to ensure that every tool is being reliably used to help resolve disagreement. Code: https://github.com/nsivaku/dart.

Tool calling has become increasingly popular for Large Language Models (LLMs). However, for large tool sets, the resulting tokens wouldexceed the LLM’s context window limit, making it impossible to include every tool. Hence, an external retriever is used to provide LLMswith the most relevant tools for a query. Existing retrieval models rank tools based on the similarity between a user query and a tool description (TD). This leads to suboptimal retrieval as user requests are often poorly aligned with the language of TD. To remedy the issue, we propose ToolDreamer, a framework that conditions retriever models to fetch tools based on hypothetical (synthetic) TD generated using an LLM, i.e., descriptions of tools that the LLM feels will be potentially useful for the query. The framework enables a more natural alignment between queries and tools within the language space of TD’s. We apply ToolDreamer on the ToolRet dataset and show that our method improves the performance of sparse and dense retrievers with and without training, showcasing its flexibility. With our proposed framework, we aim to offload a portion of the reasoning burden to the retriever so that the LLM may effectively handle a large collection of tools without inundating its context window.

Speculative decoding has emerged as a promising approach to accelerate Large Language Model inference. However, existing research has predominantly focused on 7B-70B parameters models, leaving a critical knowledge gap for small language models (1-2B parameters) that are increasingly important for edge computing and agentic AI systems. This paper presents the first comprehensive empirical study of speculative decoding techniques for small language models. We evaluate five distinct method categories across three representative model families and reveal that drafting overhead, rather than draft quality, becomes the primary bottleneck fundamentally limiting acceleration of small models. We demonstrate that traditional independent drafting fails completely due to the suboptimal architecture of available drafters, while self-drafting methods achieve meaningful acceleration only when employing sufficiently efficient draft modules. In contrast, retrieval-based methods with negligible computational overhead yield consistent gains. Based on these insights, we establish practical guidelines for effective small model acceleration.

We investigate how the choice of output format influences the performance of fine-tuned large language models on information extraction tasks. Based on over 280 experiments spanning multiple benchmarks, models and formats, we find that output formatting is a critical yet largely overlooked hyperparameter. Remarkably, in some cases, changing only the output format shifts F1 scores by over 40% despite using the same model. We further observe that no single format consistently dominates across settings, and the optimal choice depends on factors like model family and dataset characteristics. Overall, these results demonstrate that informationally equivalent output formats can produce substantial performance variation, highlighting the need to treat output formatting as a key factor in building accurate and reliable information extraction systems.

Policy-gradient reinforcement learning (PGRL) forms the backbone of current methods used to enhance alignment and reasoning in Large Language Models (LLMs). However, these methods are incompatible with diffusion based language models (dLLMs). Most attempts to apply PGRL to dLLMs, are either not scalable or use unprincipled approximations. This work, introduces PADRE a framework that uses a novel pseudo-likelihood based objective for alignment of dLLMs. Our objective has the same optima as PGRL based optimization, but does not need to evaluate exact likelihood from dLLMs. Experiments on various coding and mathematical reasoning benchmarks show that our method matches or surpasses the performance of recent dLLM training baselines such as diffu-GRPO/d1. Our approach provides a stable and practical alternative for RL-based fine-tuning of reasoning-focused dLLMs.

LLMs are powerful generators of synthetic data, which are used for training smaller, specific models. This is especially valuable for low-resource languages, where human-labelled data is scarce but LLMs can still produce high-quality text. However, LLMs differ in how useful their outputs are for training. Selecting the best LLM as a generator is challenging because extrinsic evaluation requires costly human annotations (which are often unavailable for low-resource languages), while intrinsic metrics correlate poorly with downstream performance. We introduce Round-robin Synthetic data Evaluation (RoSE), a proxy metric for selecting the best LLM generator without human test sets. RoSE trains a small model on the outputs of a candidate generator (LLM) and then evaluates it on generated synthetic examples from all other candidate LLMs. The final RoSE score is the mean performance of this small model. Across six LLMs, eleven languages, and three tasks (sentiment, topic, intent), RoSE identifies the optimal generator more often than any other intrinsic heuristics. RoSE outperforms intrinsic heuristics and comes within 0.76 percentage points of the optimal generator baseline. This result is measured in terms of downstream performance, obtained by training a small model on the chosen generator’s outputs (optimal vs. proxy-metric–selected) and evaluating it on human-labelled test data. Additionally, RoSE is the only metric to achieve a positive correlation with performance on human test data.

We investigate to what extent Multimodal Large Language Models (MLLMs) can accurately identify the orientation of input images rotated 0°, 90°, 180°, and 270°. This task demands robust visual reasoning capabilities to detect rotational cues and contextualize spatial relationships within images, regardless of their orientation. To evaluate MLLMs on these abilities, we introduce RotBench, a 350-image manually-filtered benchmark comprising lifestyle, portrait, and landscape images. Despite the relatively simple nature of this task, we show that several state-of-the-art open and proprietary MLLMs, including GPT-5, o3, and Gemini-2.5-Pro, do not reliably identify rotation in input images. Providing models with auxiliary information—including captions, depth maps, and more—or using chain-of-thought prompting offers only small and inconsistent improvements. Our results indicate that most models are able to reliably identify right-side-up (0°) images, while certain models are able to identify upside-down (180°) images. None can reliably distinguish between 90° and 270° rotated images. Simultaneously showing the image rotated in different orientations leads to moderate performance gains for reasoning models, while a modified setup using voting improves the performance of weaker models. We further show that fine-tuning does not improve models’ ability to distinguish 90° and 270° rotations, despite substantially improving the identification of 180° images. Together, these results reveal a significant gap between MLLMs’ spatial reasoning capabilities and human perception in identifying rotation.

As multilingual large language models become more widely used, ensuring their safety and fairness across diverse linguistic contexts presents unique challenges. While existing research on machine unlearning has mainly focused on monolingual settings, typically English, multilingual environments introduce additional complexities due to cross-lingual knowledge transfer and biases embedded in both pretraining and fine-tuning data. In this work, we address the problem of multilingual unlearning using the Aya-Expanse 8B model under two settings: (1) data unlearning and (2) concept unlearning. We extend benchmarks for factual knowledge and stereotypes into ten languages through translation—English, French, Arabic, Japanese, Russian, Farsi, Korean, Hindi, Hebrew, and Indonesian—spanning five language families and varying resource levels. Our experiments show that unlearning in high-resource languages tends to be more stable, with asymmetric transfer observed between typologically related languages. Moreover, analysis of linguistic distances reveals that syntactic similarity is the most predictive factor of cross-lingual unlearning effects.

Memorization has been shown to greatly inflate Large Language Models’ (LLMs) performance on domains such as math and logic, where success should primarily rely on applying generalizable reasoning rules. In many real-world applications, however, memorization is not meant to be eliminated but selectively constrained—for example, in story understanding, where background knowledge must be integrated with narrative context. Drawing on the cognitive science distinction between “verbatim” (exact recall) and “gist” (semantic abstraction) memorization, we propose a two-tier framework for analyzing how LLMs reason under different degrees of memory access. The Inductive (prompt-guided) Setting softly steers models to reason through selective, context-relevant recall, while the Restrictive Setting imposes stronger constraints by limiting verbatim memory access. Evaluating GPT-4o, LLaMA3.3-70B, and DeepSeek V3 on six character-centric story understanding benchmarks, we find up to a 45.2% accuracy drop under the Restrictive Setting, revealing strong dependence on surface recall. By contrast, the Inductive Setting maintains performance, indicating that prompting can align LLMs toward memorization-constrained reasoning.

Membership inference attacks (MIAs) reveal whether specific data was used to train machine learning models, serving as important tools for privacy auditing and compliance assessment. Recent studies have reported that MIAs perform only marginally better than random guessing against large language models, suggesting that modern pre-training approaches with massive datasets may be free from privacy leakage risks. Our work offers a complementary perspective to these findings by exploring how examining LLMs’ internal representations, rather than just their outputs, may provide additional insights into potential membership inference signals. Our framework, memTrace, follows what we call neural breadcrumbs extracting informative signals from transformer hidden states and attention patterns as they process candidate sequences. By analyzing layer-wise representation dynamics, attention distribution characteristics, and cross-layer transition patterns, we detect potential memorization fingerprints that traditional loss-based approaches may not capture. This approach yields strong membership detection across several model families achieving average AUC scores of 0.85 on popular MIA benchmarks. Our findings suggest that internal model behaviors can reveal aspects of training data exposure even when output-based signals appear protected, highlighting the need for further research into membership privacy and the development of more robust privacy-preserving training techniques for large language models.

Large language models are being rapidly applied across many fields such as healthcare, finance, transportation, and energy, among many others. These applications often involve analyzing time-series data alongside contextual information in the form of natural language to support informed decisions. However, current time-series models are limited in their ability to perform reasoning that involves both time-series and their textual content. In this work, we address this gap by introducing Chat-TS, a large language model (LLM) based framework designed to support reasoning over time series and textual data. Unlike traditional models, Chat-TS integrates time-series tokens into LLMs’ vocabulary, enhancing its reasoning ability over both modalities without compromising core natural language capabilities. To support learning and evaluation, we contribute new datasets: the TS Instruct Training Dataset (pairing diverse time-series data with relevant text instructions and responses for instruction tuning), the TS Instruct Question and Answer (QA) Gold Dataset (multiple-choice questions to evaluate multimodal reasoning), and a TS Instruct Quantitative Probing Set (a small subset of TS Instruct QA reasoning tasks alongside math and decision-making questions for LLM evaluation). We design a training strategy to preserve the inherent reasoning capabilities of LLMs while augmenting them for time-series reasoning. Experiments show that Chat-TS achieves state-of-the-art performance in multimodal reasoning tasks by maintaining strong natural language proficiency while improving time-series reasoning.

This paper investigates gender biases exhibited by LLM-based virtual assistants when providing educational recommendations, focusing on minimal gender indicators. Experimenting on Italian, a language with grammatical gender, we demonstrate that simply changing noun and adjective endings (e.g., from masculine "-o" to feminine "-a") significantly shifts recommendations. More specifically, we find that LLMs i) recommend STEM disciplines less for prompts with feminine grammatical gender and ii) narrow down the set of disciplines recommended to prompts with masculine grammatical gender; these effects persist across multiple commercial LLMs (from OpenAI, Anthropic, and Google). We show that grammatical gender cues alone trigger substantial distributional shifts in educational recommendations, and up to 76% of the bias exhibited when using prompts with proper names is already present with grammatical gender markers alone.Our findings highlight the need for robust bias evaluation and mitigation strategies before deploying LLM-based virtual assistants in student-facing contexts and the risks of using general purpose LLMs for educational applications, especially in languages with grammatical gender.

Enterprise documents, such as forms and reports, embed critical information for downstream applications like data archiving, automated workflows, and analytics. Although generalist Vision Language Models (VLMs) perform well on established document understanding benchmarks, their ability to conduct holistic, fine-grained structured extraction across diverse document types and flexible schemas is not well studied. Existing Key Entity Extraction (KEE), Relation Extraction (RE), and Visual Question Answering (VQA) datasets are limited by narrow entity ontologies, simple queries, or homogeneous document types, often overlooking the need for adaptable and structured extraction. To address these gaps, we introduce ExStrucTiny, a new benchmark dataset for structured Information Extraction (IE) from document images, unifying aspects of KEE, RE, and VQA. Built through a novel pipeline combining manual and synthetic human-validated samples, ExStrucTiny covers more varied document types and extraction scenarios. We analyze open and closed VLMs on this benchmark, highlighting challenges such as schema adaptation, query under-specification, and answer localization. We hope our work provides a bedrock for improving generalist models for structured IE in documents.

Despite the impressive performance of vision-language models (VLMs) on downstream tasks, their ability to understand and reason about causal relationships in visual inputs remains unclear. Robust causal reasoning is fundamental to solving complex high-level reasoning tasks, yet existing benchmarks often include a mixture of reasoning questions, and VLMs can frequently exploit object recognition and activity identification as shortcuts to arrive at the correct answers, making it challenging to truly assess their causal reasoning abilities. To bridge this gap, we introduce VQA-Causal and VCR-Causal, two new benchmarks specifically designed to isolate and rigorously evaluate VLMs’ causal reasoning abilities. Our findings reveal that while VLMs excel in object and activity recognition, they perform poorly on causal reasoning tasks, often only marginally surpassing random guessing. Further analysis suggests that this limitation stems from a severe lack of causal expressions in widely used training datasets, where causal relationships are rarely explicitly conveyed. We additionally explore fine-tuning strategies with hard negative cases, showing that targeted fine-tuning can improve model’s causal reasoning while maintaining generalization and downstream performance. Our study highlights a key gap in current VLMs and lays the groundwork for future work on causal understanding. We will release the code upon acceptance.

Multimodal Large Language Models (MLLMs) show progress across many visual–language tasks; however, their capacity to evaluate artistic expression remains limited: aesthetic concepts are inherently abstract and open-ended, and multimodal artwork annotations are scarce. We introduce KidsArtBench, a new benchmark of over 1k children’s artworks (ages 5-15) annotated by 12 expert educators across 9 rubric-aligned dimensions, together with expert comments for feedback. Unlike prior aesthetic datasets that provide single scalar scores on adult imagery, KidsArtBench targets children’s artwork and pairs multi-dimensional annotations with comment supervision to enable both ordinal assessment and formative feedback. Building on this resource, we propose an attribute-specific multi-LoRA approach – where each attribute corresponds to a distinct evaluation dimension (e.g., Realism, Imagination) in the scoring rubric – with Regression-Aware Fine-Tuning (RAFT) to align predictions with ordinal scales. On Qwen2.5-VL-7B, our method increases correlation from 0.468 to 0.653, with the largest gains on perceptual dimensions and narrowed gaps on higher-order attributes. Our results show that educator-aligned supervision and attribute-aware training yield pedagogically meaningful evaluations and establish a rigorous testbed for sustained progress in educational AI. We release data and code with ethics documentation.

Model steering, which involves intervening on hidden representations at inference time, has emerged as a lightweight alternative to finetuning for precisely controlling large language models. While steering efficacy has been widely studied, evaluations of whether interventions alter *only* the intended property remain limited, especially with respect to unintended changes in behaviors related to the target property. We call this notion specificity. We propose a framework that distinguishes three dimensions of specificity: general (preserving fluency and unrelated abilities), control (preserving related control properties), and robustness (preserving control properties under distribution shifts). We study two safety-critical use cases: steering models to reduce overrefusal and faithfulness hallucinations, and show that while steering achieves high efficacy and largely maintains general and control specificity, it consistently fails to preserve robustness specificity. In the case of overrefusal steering, for example, all steering methods reduce overrefusal without harming general abilities and refusal on harmful queries; however, they substantially increase vulnerability to jailbreaks. Our work provides the first systematic evaluation of specificity in model steering, showing that standard efficacy and specificity checks are insufficient, because without robustness evaluation, steering methods may appear reliable even when they compromise model safety.

Multilingual domain adaptation (ML-DA) enables large language models (LLMs) to acquire domain knowledge across languages. Despite many methods, how domain knowledge is acquired within a language and transferred across languages remains, leading to suboptimal performance, particularly in low-resource settings.This work examines the learning dynamics of LLMs during ML-DA. Because prior ML-DA studies often train and evaluate on datasets with mismatched knowledge coverage, we propose AdaXEval, an adaptive evaluation method that constructs multiple-choice QA datasets from the same bilingual domain corpus used for training, thereby enabling direct analysis of multilingual knowledge acquisition.Through continual training of LLMs with diverse data recipes, we track how LLMs acquire domain facts and pinpoint the loss shielding mechanism behind the knowledge memorization and generalization in domain adaptation. Our experiments on multilingual LLMs reveal that cross-lingual transfer remains challenging.The code is released.

Tokenization is a critical component of language model pretraining, yet standard tokenization methods often prioritize information-theoretical goals like high compression and low fertility rather than linguistic goals like morphological alignment. In fact, they have been shown to be suboptimal for morphologically rich languages, where tokenization quality directly impacts downstream performance. In this work, we investigate morphologically-aware tokenization for Latin, a morphologically rich language that is medium-resource in terms of pretraining data, but high-resource in terms of curated lexical resources – a distinction that is often overlooked but critical in discussions of low-resource language modeling. We find that morphologically-guided tokenization improves overall performance on four downstream tasks. Performance gains are most pronounced for out of domain texts, highlighting our models’ improved generalization ability. Our findings demonstrate the utility of linguistic resources to improve language modeling for morphologically complex languages. For low-resource languages that lack large-scale pretraining data, the development and incorporation of linguistic resources can serve as a feasible alternative to improve LM performance.

Curriculum learning—organizing training data from easy to hard—has improved efficiency across machine learning domains, yet remains underexplored for language model pretraining. We present the first systematic investigation of curriculum learning in LLM pretraining, with over 200 models trained on up to 100B tokens across three strategies: vanilla curriculum learning, pacing-based sampling, and interleaved curricula, guided by six difficulty metrics spanning linguistic and information-theoretic properties. We evaluate performance on eight benchmarks under three realistic scenarios: limited data, unlimited data, and continual training. Our experiments show that curriculum learning consistently accelerates convergence in early and mid-training phases, reducing training steps by 18-45% to reach baseline performance. When applied as a warmup strategy before standard random sampling, curriculum learning yields sustained improvements up to 3.5%. We identify compression ratio, lexical diversity (MTLD), and readability (Flesch Reading Ease) as the most effective difficulty signals. Our findings demonstrate that data ordering—orthogonal to existing data selection methods—provides a practical mechanism for more efficient LLM pretraining.

Recent advances in multimodal large language models (MLLMs) offer a promising approach for natural language-based scene change queries in virtual reality (VR). Prior work on applying MLLMs for object state understanding has focused on egocentric videos that capture the camera wearer’s interactions with objects. However, object state changes may occur in the background without direct user interaction, lacking explicit motion cues and making them difficult to detect. Moreover, no benchmark exists for evaluating this challenging scenario. To address these challenges, we introduce ObjChangeVR-Dataset, specifically for benchmarking the question-answering task of object state change. We also propose ObjChangeVR, a framework that combines viewpoint-aware and temporal-based retrieval to identify relevant frames, along with cross-view reasoning that reconciles inconsistent evidence from multiple viewpoints. Extensive experiments demonstrate that ObjChangeVR significantly outperforms baseline approaches across multiple MLLMs.

A common solution for mitigating outdated or incorrect information in Large Language Models (LLMs) is to provide updated facts in-context or through knowledge editing. However, these methods introduce knowledge conflicts when the knowledge update fails to overwrite the model’s parametric knowledge, which propagate to faulty reasoning. Current benchmarks for this problem, however, largely focus only on single knowledge updates and fact recall without evaluating how these updates affect downstream reasoning. In this work, we introduce Tʀᴀᴄᴋ (*Testing Reasoning Amid Conflicting Knowledge*), a new benchmark for studying how LLMs propagate new knowledge through multi-step reasoning when it conflicts with the model’s initial parametric knowledge. Spanning three reasoning-intensive scenarios (WIKI, CODE, and MATH), Tʀᴀᴄᴋ introduces multiple, realistic conflicts to mirror real-world complexity. Our results on Tʀᴀᴄᴋ reveal that providing updated facts to models for reasoning can worsen performance compared to providing no updated facts to a model, and that this performance degradation exacerbates as more updated facts are provided. We show this failure stems from both inability to faithfully integrate updated facts, but also flawed reasoning even when knowledge is integrated. Tʀᴀᴄᴋ provides a rigorous new benchmark to measure and guide future progress on propagating conflicting knowledge in multi-step reasoning.

Understanding emotion from speech requires sensitivity to both lexical and acoustic cues. However, it remains unclear whether large audio language models (LALMs) genuinely process acoustic information or rely primarily on lexical contents. We present LISTEN (Lexical vs. Acoustic Speech Test for Emotion in Narratives), a controlled benchmark designed to disentangle lexical reliance from acoustic sensitivity in emotion understanding. Across evaluations of six state-of-the-art LALMs, we observe a consistent lexical dominance. Models predict “neutral” when lexical cues are neutral or absent, show limited gains under cue alignment, and fail to classify distinct emotions under cue conflict. In paralinguistic settings, performance approaches chance. These results indicate that current LALMs largely “transcribe” rather than “listen,” relying heavily on lexical semantics while underutilizing acoustic cues. LISTEN offers a principled framework for assessing emotion understanding in multimodal models.

Recent advances in self-supervised learning (SSL) have led to powerful speech representation models, yet their robustness in real-world conversational settings remains largely untested. Most existing benchmarks focus on clean, single-speaker, single-channel audio, failing to reflect the complexities of natural human interaction—where background noise, reverberation, and overlapping speech are the norm. To bridge these critical gaps, we present the Conversational Speech Processing Benchmark (CSPB), a new benchmark designed to assess the robustness of SSL speech models in realistic conversational scenarios. CSPB is constructed from four multi-party datasets—AMI, AliMeeting, MMCSG, and DiPCo—and supports both single-channel and multi-channel evaluation. By releasing CSPB as an open-source toolkit, we aim to establish a unified framework for evaluating and advancing robust, spatially-aware self-supervised speech models.

Text anonymization is a critical task for enabling research and development in high-stakes domains containing private data, like medicine, law, and social services. While much research has focused on redacting sensitive content from text, substantially less work has focused on what to replace redacted content with, which can enhance privacy and becomes increasingly important with greater levels of redaction. In this work, we formulate predicting replacements for sensitive spans as a research task with principled use-inspired evaluation criteria. We further propose a multi-token completion method for accomplishing this task that is designed to preserve consistency with low compute requirements, thus facilitating practitioners to anonymize data locally before sharing it externally. Human and automated annotations demonstrate that our approach produces more realistic text and better preserves utility than alternative infilling methods and differentially private mechanisms across multiple domains without retraining. Overall, our work explores the under-studied task of what to replace redacted content with and contributes grounded evaluations capturing utility, facilitating future work.

Large language models (LLMs) excel in English but still struggle with complex reasoning in many low-resource languages (LRLs). Existing methods align LLMs with multilingual encoders, such as LangBridge and MindMerger, raising the accuracy for mid and high-resource languages, yet large performance gap remains for LRLs. We present MERLIN, a model-stacking framework that iteratively refines in 2-stages based on a curriculum strategy (from general to specific where general is bilingual bitext and specific is task-specific data) and adapts only a small set of DoRA weights. On the AfriMGSM benchmark MERLIN improves exact-match accuracy by +12.9 pp over MindMerger and outperforms GPT-4o-mini by 15.2 pp. It also yields consistent gains on MGSM and MSVAMP (+0.9 and +2.8 pp), demonstrating effectiveness across both low and high-resource settings.

Large audio-language models increasingly operate on raw speech inputs, enabling more seamless integration across domains such as voice assistants, education, and clinical triage. This transition, however, introduces a distinct class of vulnerabilities that remain largely uncharacterized. We examine the security implications of this modality shift by designing a text-to-audio jailbreak that embeds disallowed directives within a narrative-style audio stream. The attack leverages an advanced instruction-following text-to-speech (TTS) model to exploit structural and acoustic properties, thereby circumventing safety mechanisms primarily calibrated for text. When delivered through synthetic speech, the narrative format elicits restricted outputs from state-of-the-art models, including Gemini 2.0 Flash, achieving a 98.26% success rate that substantially exceeds text-only baselines. These results highlight the need for safety frameworks that jointly reason over linguistic and paralinguistic representations, particularly as speech-based interfaces become more prevalent.

Sparse autoencoders (SAEs) are commonly used to interpret the internal activations of large language models (LLMs) by mapping them to human-interpretable concept representations. While existing evaluations of SAEs focus on metrics such as the reconstruction-sparsity tradeoff, human (auto-)interpretability, and feature disentanglement, they overlook a critical aspect: the robustness of concept representations to input perturbations. We argue that robustness must be a fundamental consideration for concept representations, reflecting the fidelity of concept labeling. To this end, we formulate robustness quantification as input-space optimization problems and develop a comprehensive evaluation framework featuring realistic scenarios in which adversarial perturbations are crafted to manipulate SAE representations. Empirically, we find that tiny adversarial input perturbations can effectively manipulate concept-based interpretations in most scenarios without notably affecting the base LLM’s activations. Overall, our results suggest that SAE concept representations are fragile and without further denoising or postprocessing they might be ill-suited for applications in model monitoring and oversight.

Leveraging a dataset of paired narratives, we investigate the extent to which large language models (LLMs) can reliably separate incoherent and coherent stories.A probing study finds that LLMs’ internal representations can reliably identify incoherent events in narratives. However, this separation disappears by the narrative’s end, and weakens when the differences between coherent and incoherent stories are more subtle. When asked to rate overall coherence of narratives after reading, LLMs generate responses that fail to satisfactorily separate the coherent and incoherent narratives.Reasoning models tested do not eliminate these deficits, indicating that thought strings may not be able to fully address the discrepancy between model internal state and behavior.Additionally, we find that LLMs appear to be more sensitive to incoherence resulting from an event that violates the setting (e.g., a rainy day in the desert) than to incoherence arising from a character violating an established trait (e.g., Mary, a vegetarian, later orders a cheeseburger), suggesting that LLMs may rely more on prototypical world knowledge than building coherence through a meaning-based world model of the narrative setting. Together, our results indicate that LLMs lack robustness in their ability to recognize incoherence in narratives.

Working memory, or the ability to hold and manipulate information in the mind, is a critical component of human intelligence and executive functioning. It is correlated with performance on various cognitive tasks, including measures of fluid intelligence, which encompasses reasoning and problem solving. We use a comprehensive set of classic working memory tasks to estimate the working memory capacity of large language models (LLMs). We find that in most cases, LLMs exceed normative human scores. However, we do not find that the increased capacity of working memory is associated with higher performance on other executive functioning tasks or problem solving benchmarks. These results suggest that LLMs may have deficits in attentional control and cognitive flexibility, which result in difficulties with inhibiting automatic responses and adapting to shifting information. Our findings suggest that reasoning models, although they often do not currently fully compensate for these deficits, may have the potential to do so in the future.

Language models (LMs) have been reported to implicitly encode character-level information, despite not being explicitly provided during training. However, the mechanisms underlying this phenomenon remain largely unexplored. To reveal the mechanisms, we analyze how models acquire character-level knowledge by comparing LMs trained under controlled settings, such as specifying the pre-training dataset or tokenizer, with those trained under standard settings. We categorize the contributing factors into those independent of tokenization. Our analysis reveals that merge rules and orthographic constraints constitute primary factors arising from tokenization, whereas semantic associations of substrings and syntactic information function as key factors independent of tokenization.

Turn-taking is a fundamental component of human communication and is signalled through complex cues distributed across lexical, temporal, and prosodic information. Full-duplex models of spoken dialogue integrate these information sources to produce impressive turn-taking behaviour. Yet, existing evaluations of their turn-taking capabilities do not address which information sources drive predictions.We present a systematic analysis of the role of lexical-temporal features on the predictability of turn structure by examining PairwiseTurnGPT, a full-duplex model of spoken dialogue transcripts. Through PCA, mixed-effects modelling, and temporal surprisal analysis, we reveal context-dependent patterns: linguistic fluency paradoxically creates overconfidence at intermediate completion points, while turn-shift overlap dominates boundary detection. Our findings uncover where lexical-temporal information suffices and where additional cues become necessary, establishing a deeper understanding of how turn-taking cues are distributed and how to evaluate dialogue systems.

Evaluating the quality of children’s utterances in adult-child dialogue remains challenging due to insufficient context-sensitive metrics. Common proxies such as Mean Length of Utterance (MLU), lexical diversity (vocd-D), and readability indices (Flesch-Kincaid Grade Level, Gunning Fog Index) are dominated by length and ignore conversational context, missing aspects of response quality such as reasoning depth, topic maintenance, and discourse planning. We introduce an LLM-as-a-judge framework that first classifies the Previous Adult Utterance Type and then scores the child’s response along two axes: Expansion (contextual elaboration and inferential depth) and Independence (the child’s contribution to advancing the discourse). These axes reflect fundamental dimensions in child language development, where Expansion captures elaboration, clause combining, and causal and contrastive connectives. Independence captures initiative, topic control, decreasing reliance on adult scaffolding through growing self-regulation, and audience design. We establish developmental validity by showing age-related patterns and demonstrate predictive value by improving age estimation over common baselines. We further confirm semantic sensitivity by detecting differences tied to discourse relations. Our metrics align with human judgments, enabling large-scale evaluation. This shifts child utterance assessment from simply measuring length to evaluating how meaningfully the child’s speech contributes to and advances the conversation within its context.

Ancient people translated classical Chinese into Japanese using a system of annotations placed around characters. We abstract this process as sequence tagging tasks and fit them into modern language technologies. The research on this annotation and translation system faces a low resource problem. We alleviate this problem by introducing an LLM-based annotation pipeline and constructing a new dataset from digitized open-source translation data. We show that in the low-resource setting, introducing auxiliary Chinese NLP tasks enhances the training of sequence tagging tasks. We also evaluate the performance of Large Language Models (LLMs) on this task. While they achieve high scores on direct machine translation, our method could serve as a supplement to LLMs to improve the quality of character’s annotation.

Large Audio-Language Models (LALMs) are often constrained by short audio context windows, even when their text backbones support long contexts, limiting long-form audio understanding. Prior work has introduced context-extension methods (e.g. YaRN) on unimodal LLMs, yet their application to LALMs remains unexplored. First, building on RoPE-based context extension, we introduce Partial YaRN, a training-free, modality-decoupled extension method that modifies only audio token positions, leaving text positions intact to preserve the base LLM’s text capabilities. Second, we propose Virtual Longform Audio Training (VLAT), a training strategy that extends Partial YaRN into a training-time positional augmentation. VLAT simulates diverse audio lengths during training, enabling generalization to inputs far longer than those seen in training. Our experiments on SALMONN and Qwen2-Audio confirm that Partial YaRN outperforms the original models across wide range of settings, and VLAT provides substantial performance improvement on long audio of unseen lengths.

Hallucinations remain a persistent challenge for vision–language models (VLMs), which often describe nonexistent objects or fabricate facts. Existing detection methods typically operate after text generation, making intervention both costly and untimely. We investigate whether hallucination risk can instead be predicted before any token is generated by probing a model’s internal representations in a single forward pass. Across a diverse set of vision–language tasks and eight modern VLMs, including Llama-3.2-Vision, Gemma-3, Phi-4-VL, and Qwen2.5-VL, we examine three families of internal representations: (i) visual-only features without multimodal fusion, (ii) vision token representations within the text decoder, and (iii) query-token representations that integrate visual and textual information before generation. Probes trained on these representations achieve strong hallucination-detection performance without decoding, reaching up to 0.93 AUROC on Gemma-3-12B, Phi-4-VL 5.6B, and Molmo 7B. Late query-token states are the most predictive for most models, while visual or mid layer features dominate in a few architectures (e.g., ∼0.79 AUROC for Qwen2.5-VL-7B using visual-only features). These results demonstrate that (1) hallucination risk is detectable pre-generation, (2) the most informative layer and modality vary across architectures, and (3) lightweight probes has the potential to enable early abstention, selective routing, and adaptive decoding to improve both safety and efficiency.

Large language models remain predominantly English-centric, which limits their utility for underrepresented languages. We help bridge this gap for Hindi with Llama-3-Nanda-10B-Chat (aka Nanda-10B) and Llama-3.1-Nanda-87B-Chat (aka Nanda-87B), forming the Nanda family of open-weight bilingual models (https://github.com/MBZUAI-IFM/Nanda-Family). Our approach integrates: (i) a tokenizer extending Llama’s vocabulary with 20% Hindi-specific tokens, thus halving Hindi tokenization fertility while preserving English efficiency, (ii) Hindi-first parameter-efficient continual pretraining using Llama Pro on a 65B-token corpus spanning Devanagari script, code-mixed, and Romanized Hindi, and (iii) bilingual instruction and safety alignment on a large culturally grounded dataset. The resulting Nanda models outperform open-weight LLMs of comparable size: Nanda-87B yields high generative quality, and Nanda-10B shows competitive general-purpose performance. Nanda-87B demonstrates state-of-the-art performance on summarization, translation, transliteration, and instruction following. Moreover, both models achieve state-of-the-art performance in safety and in cultural knowledge. Our results demonstrate that careful tokenizer design, data curation, and continual pretraining can yield capable and safe LLMs for resource-poor languages without compromising English performance.

The role of world knowledge has been particularly crucial to predict the discourse connective that marks the discourse relation between two arguments, with language models (LMs) being generally successful at this task. We flip this premise in our work, and instead study the inverse problem of understanding whether discourse connectives can inform LMs about the world. To this end, we present Wugnectives, a dataset of 8,880 stimuli that evaluates LMs’ inferences about novel entities in contexts where connectives link the entities to particular attributes. On investigating 17 different LMs at various scales, and training regimens, we found that tuning an LM to show reasoning behavior yields noteworthy improvements on most connectives. At the same time, there was a large variation in LMs’ overall performance across connective type, with all models systematically struggling on connectives that express a concessive meaning. Our findings pave the way for more nuanced investigations into the functional role of language cues as captured by LMs.We release Wugnectives at https://github.com/kanishkamisra/wugnectives

Existing multilingual long-context benchmarks, often based on the popular needle-in-a-haystack test, primarily evaluate a model’s ability to locate specific information buried within irrelevant texts. However, such a retrieval-centric approach is myopic and inherently limited, as successful recall alone does not indicate a model’s capacity to reason over extended contexts. Moreover, these benchmarks are susceptible to data leakage, short-circuiting, and risk making the evaluation a priori identifiable. To address these limitations, we introduce MLRBench, a new synthetic benchmark for multilingual long-context reasoning. Unlike existing benchmarks, MLRBench goes beyond surface-level retrieval by including bAbI-style tasks that test multi-hop inference, aggregation, and epistemic reasoning. Spanning seven languages, we design MLRBench to be parallel, resistant to leakage, and scalable to arbitrary context lengths. Our extensive experiments with an open-weight large language model (LLM) reveal a pronounced gap between high- and low-resource languages, particularly for tasks requiring the model to aggregate multiple facts or predict the absence of information. We also find that, in multilingual settings, LLMs effectively utilize less than 30% of their claimed context length. Although off-the-shelf Retrieval-Augmented Generation helps alleviate this to a certain extent, it does not solve the long-context problem. We open-source MLRBench to enable future research in the improved evaluation and training of multilingual LLMs.

Current evaluation of large language models (LLMs) overwhelmingly prioritizes accuracy; however, in real-world and safety-critical applications, the ability to abstain when uncertain is equally vital for trustworthy deployment. We introduce a unified benchmark and evaluation protocol for abstention in medical multiple-choice question answering (MCQA), integrating conformal prediction, adversarial question perturbations, and explicit abstention options. Our systematic evaluation of both open- and closed-source LLMs reveals that even state-of-the-art, high-accuracy models often fail to abstain when uncertain. Notably, providing explicit abstention options consistently increases model uncertainty and safer abstention, far more than input perturbations, while scaling model size or advanced prompting brings little improvement. These findings highlight the central role of abstention mechanisms for trustworthy LLM deployment and offer practical guidance for improving safety in high-stakes applications.

Artificial intelligence (AI) and large language models (LLMs) in healthcare require advanced clinical skills (CS), yet current benchmarks fail to evaluate these comprehensively. We introduce MedQA-CS, an AI-SCE framework inspired by medical education’s Objective Structured Clinical Examinations (OSCEs), to address this gap. MedQA-CS evaluates LLMs through two instruction-following tasks—LLM-as-medical-student and LLM-as-CS-examiner—designed to reflect real clinical scenarios. Our contributions include developing MedQA-CS, a comprehensive evaluation framework with publicly available data and expert annotations, and providing the quantitative and qualitative assessment of LLMs as reliable judges in CS evaluation. Our experiments show that MedQA-CS is a more challenging benchmark for evaluating clinical skills than traditional multiple-choice QA benchmarks (e.g., MedQA). Combined with existing benchmarks, MedQA-CS enables a more comprehensive evaluation of LLMs’ clinical capabilities for both open- and closed-source LLMs.

Modelling a language model for a multi-lingual scenario includes several potential challenges, among which catastrophic forgetting is the major challenge. For example, small language models (SLM) built for low-resource languages by adapting large language models (LLMs) pose the challenge of catastrophic forgetting. This work proposes to employ a continual learning strategy using parts-of-speech (POS)-based code-switching along with a replay adapter strategy to mitigate the identified gap of catastrophic forgetting while training SLM from LLM. Experiments conducted on vision language tasks such as visual question answering and language modelling task exhibits the success of the proposed architecture.

Multi-domain image-to-image translation requires grounding semantic differences expressed in natural language prompts into corresponding visual transformations, while preserving unrelated structural and semantic content. Existing methods struggle to maintain structural integrity and provide fine-grained, attribute-specific control, especially when multiple domains are involved. We propose LACE (Language-grounded Attribute-Controllable Translation), built on two components: (1) a GLIP-Adapter that fuses global semantics with local structural features to preserve consistency, and (2) a Multi-Domain Control Guidance mechanism that explicitly grounds the semantic delta between source and target prompts into per-attribute translation vectors, aligning linguistic semantics with domain-level visual changes. Together, these modules enable compositional multi-domain control with independent strength modulation for each attribute. Experiments on CelebA(Dialog) and BDD100K demonstrate that LACE achieves high visual fidelity, structural preservation, and interpretable domain-specific control, surpassing prior baselines. This positions LACE as a cross-modal content generation framework bridging language semantics and controllable visual translation. Code will be publicly available.

Large language models (LLMs) encode rich cultural knowledge learned from diverse web-scale data, offering an unprecedented opportunity to model cultural commonsense at scale. Yet this knowledge remains mostly implicit and unstructured, limiting its interpretability and use. We present an iterative, prompt-based framework for constructing a Cultural Commonsense Knowledge Graph (CCKG) that treats LLMs as cultural archives, systematically eliciting culture-specific entities, relations, and practices and composing them into multi-step inferential chains across languages. We evaluate CCKG on five countries with human judgments of cultural relevance, correctness, and path coherence. We find that the cultural knowledge graphs are better realized in English, even when the target culture is non-English (e.g., Chinese, Indonesian, Arabic), indicating uneven cultural encoding in current LLMs. Augmenting smaller LLMs with CCKG improves performance on cultural reasoning and story generation, with the largest gains from English chains. Our results show both the promise and limits of LLMs as cultural technologies and that chain-structured cultural knowledge is a practical substrate for culturally grounded NLP.

Grammar comprehension is a critical capability for large language models (LLMs) to achieve fluency in a target language. In low-resource settings, such as the case with Arabic, limited availability of high-quality data can lead to significant gaps in grammatical understanding, making systematic evaluation essential. We introduce Nahw, a comprehensive benchmark for Arabic grammar that covers both theoretical knowledge and practical applications, including grammatical error detection, correction, and explanation. We evaluate a range of LLMs on these tasks and find that many models still exhibit substantial deficiencies in Arabic grammar comprehension, with GPT-4o achieving a score of 67% on average over all tasks, while the best performing Arabic model in our experiment (ALLaM-7B) achieving 42%. Our experiments also demonstrate that while fine-tuning with synthetic data can improve performance, it does not match the effectiveness of training on natural, high-quality data.

We examine, analyze, and compare four representative creativity measures—perplexity, LLM-as-a-Judge, the Creativity Index (CI; measuring n-gram overlap with web corpora), and syntactic templates (detecting repetition of common part-of-speech patterns)—across the diverse creative domains, such as creative writing, unconventional problem-solving, and research ideation. For each domain, we compile datasets with human-aligned creative and uncreative examples and evaluate each metric’s ability to discriminate between the two sets. Our analyses reveal limited consistency both across domains and metrics, as metrics that distinguish creativity in one domain fail in others (e.g., CI correctly distinguishes in creative writing but fails in problem-solving), and different metrics often disagree on the same data points (e.g., CI suggests one set to be more creative, while perplexity indicates the other set to be more creative.) We highlight key limitations, such as perplexity reflecting fluency rather than novelty; LLM-as-a-Judge producing inconsistent judgments under minor prompt variations and exhibiting bias towards particular labels; CI primarily measuring lexical diversity, with high sensitivity to implementation choices; and syntactic templates being ineffective in settings dominated by formulaic language. Our findings underscore the need for more robust, generalizable evaluation frameworks that better align with human judgments of creativity. We release the datasets and evaluation code: https://github.com/lichun-19/creative_eval.

Building effective tokenizers for multilingual Large Language Models (LLMs) requires careful control over language-specific data mixtures. While a tokenizer’s compression performance critically affects the efficiency of LLM training and inference, existing approaches rely on heuristics or costly large-scale searches to determine optimal language ratios. We introduce Tokenizer Regression for Optimal Data MiXture (TReX), a regression-based framework that efficiently predicts the optimal data mixture for tokenizer training. TReX trains small-scale proxy tokenizers on random mixtures, gathers their compression statistics, and learns to predict compression performance from data mixtures. This learned model enables scalable mixture search before large-scale tokenizer training, mitigating the accuracy-cost trade-off in multilingual tokenizer design. Tokenizers trained with TReX’s predicted mixtures outperform mixtures based on LLaMA3 and uniform distributions by up to 12% in both in- and out-of-distribution compression efficiency, demonstrating strong scalability, robustness, and practical effectiveness.

Naive joint training of large language models (LLMs) for multilingual preference alignment can suffer from negative interference. This is a known issue in multilingual training, where conflicting objectives degrade overall performance. However, the impact of this phenomenon in the context of multilingual preference alignment remains largely underexplored. To address this issue, we propose ConGrad, an effective and scalable filtering method that mitigates this interference by identifying and selecting preference samples that exhibit high cross-lingual affinity. Based on principles of multi-objective optimization, our approach computes an aggregated, cross-lingually beneficial gradient direction and uses this to filter for samples whose individual gradients align with this consensus direction. To ensure scalability for LLMs, we incorporate a sublinear gradient compression strategy that reduces memory overhead during gradient accumulation. We integrate ConGrad into a self-rewarding framework and evaluate on LLaMA3-8B and Gemma2-2B across 10 languages. Results show that ConGrad consistently outperforms strong baselines in both seen and unseen languages, with minimal alignment tax.

Large Language Models exhibit implicit personalities in their generation, but reliably controlling or aligning these traits to meet specific needs remains an open challenge. The need for effective mechanisms for behavioural manipulation of the model during generation is a critical gap in the literature that needs to be fulfilled. Personality-aware LLMs hold a promising direction towards this objective.However, the relationship between these psychological constructs and their representations within LLMs remains underexplored and requires further investigation. Moreover, it is intriguing to understand and study the use of these representations to steer the models’ behaviour. We propose a novel pipeline that extracts hidden state activations from transformer layers using the Big Five Personality Traits (Openness, Conscientiousness, Extraversion, Agreeableness and Neuroticism), which is a comprehensive and empirically validated framework to model human personality applies low-rank subspace discovery methods, andidentifies trait-specific optimal layers across different model architectures for robust injection. The resulting personality-aligned directions are then operationalised through a flexible steering framework with dynamic layer selection, enabling precise control of trait expression in LLM outputs. Our findings reveal that personality traits occupy a low-rank shared subspace, and that these latent structures can be transformed into actionable mechanisms for effective steering through careful perturbations without impacting the fluency, variance and general capabilities, helping to bridge the gap between psychological theory and practical model alignment.

Speculative generation has emerged as a promising technique to accelerate inference in large language models (LLMs) by leveraging parallelism to verify multiple draft tokens simultaneously. However, the fundamental limits on the achievable speedup remain poorly understood. In this work, we establish the first “tight” lower bounds on the runtime of any deterministic speculative generation algorithm. This is achieved by drawing a parallel between the token generation process and branching random walks, which allows us to analyze the optimal draft tree selection problem. We prove, under basic assumptions, that the expected number of tokens successfully predicted per speculative iteration is bounded as \mathbb{E}[X] ≤ (𝜇 + 𝜇₍₂₎)log(B )/𝜇² + O(1), where B is the verifier’s batch size, 𝜇 is the expected entropy of the verifier’s output distribution, and 𝜇₍₂₎ is this entropy’s second moment. This result provides new insights into the limits of parallel token generation, and could guide the design of future speculative decoding systems. Empirical evaluations on Llama models validate our theoretical predictions, confirming the tightness of our bounds in practical settings.

Claim verification is a core module in automated fact-checking systems, tasked with determining claim veracity using retrieved evidence. This work presents KG-CRAFT, a novel knowledge graph-based contrastive reasoning method that enhances automatic claim verification by LLMs. Our approach first constructs a knowledge graph from claims and associated reports, then formulates contextually relevant contrastive questions based on the knowledge graph structure. These questions guide the distillation of evidence-based reports, which are synthesised into a concise summary for veracity assessment. Extensive evaluations on two real-world datasets (LIAR-RAW and RAWFC) demonstrate that our method achieves a new state-of-the-art in predictive performance. Comprehensive analyses validate in detail the effectiveness of our knowledge graph-based contrastive reasoning approach in improving LLMs’ fact-checking capabilities.

The accelerating growth of scientific publications has intensified the need for scalable, trustworthy systems to synthesize knowledge across diverse literature. While recent retrieval-augmented generation (RAG) methods have improved access to scientific information, they often overlook citation graph structure, adapt poorly to complex queries, and yield fragmented, hard-to-verify syntheses. We introduce SciRAG, an open-source framework for scientific literature exploration that addresses these gaps through three key innovations: (1) adaptive retrieval that flexibly alternates between sequential and parallel evidence gathering; (2) citation-aware symbolic reasoning that leverages citation graphs to organize and filter supporting documents; and (3) outline-guided synthesis that plans, critiques, and refines answers to ensure coherence and transparent attribution. Extensive experiments across multiple benchmarks such as QASA and ScholarQA demonstrate that SciRAG outperforms prior systems in factual accuracy and synthesis quality, establishing a new foundation for reliable, large-scale scientific knowledge aggregation.

Fine-tuning Large Language Models (LLMs) on sensitive datasets carries a substantial risk of unintended memorization and leakage of Personally Identifiable Information (PII), which can violate privacy regulations and compromise individual safety. In this work, we systematically investigate a critical and underexplored vulnerability: the exposure of PII that appears only in model inputs, not in training targets. Using both synthetic and real-world datasets, we design controlled extraction probes to quantify unintended PII memorization and study how factors such as language, PII frequency, task type, and model size influence memorization behavior. We further benchmark four privacy-preserving approaches including differential privacy, machine unlearning, regularization, and preference alignment, evaluating their trade-offs between privacy and task performance. Our results show that post-training methods generally provide more consistent privacy–utility trade-offs, while differential privacy achieves strong reduction in leakage in specific settings, although it can introduce training instability. These findings highlight the persistent challenge of memorization in fine-tuned LLMs and emphasize the need for robust, scalable privacy-preserving techniques.

People increasingly rely on Large Language Models (LLMs) for moral advice, which may influence humans’ decisions. Yet, little is known about how closely LLMs align with human moral judgments. To address this, we introduce the Moral Dilemma Dataset, a benchmark of 1,618 real-world moral dilemmas paired with a distribution of human moral judgments consisting of a binary evaluation and a free-text rationale. We treat this problem as a pluralistic distributional alignment task, comparing the distributions of LLM and human judgments across dilemmas. We find that models reproduce human judgments only under high consensus; alignment deteriorates sharply when human disagreement increases. In parallel, using a 60-value taxonomy built from 3,783 value expressions extracted from rationales, we show that LLMs rely on a narrower set of moral values than humans. These findings reveal a pluralistic moral gap–a mismatch in both the distribution and diversity of values expressed. To close this gap, we introduce Dynamic Moral Profiling (DMP), a Dirichlet-based sampling method that conditions model outputs on human-derived value profiles. DMP improves alignment by 64.3% and enhances value diversity, offering a step toward more pluralistic and human-aligned moral guidance.

Existing datasets for multimodal table understanding, such as MMTab, primarily provide short factual answers without explicit multi-step reasoning supervision. Models trained on these datasets often generate brief responses that offers insufficient accuracy and limited interpretability into how these models arrive at the final answer. We introduce CoReTab, a code-driven reasoning framework that produces scalable, interpretable, and automatically verifiable annotations by coupling multi-step reasoning with executable Python code. Using the CoReTab framework, we curate a dataset of 115K verified samples averaging 529 tokens per response and fine-tune open-source MLLMs through a three-stage pipeline. We evaluate the resulting model trained on CoReTab across 17 MMTab benchmarks spanning table question answering, fact verification, and table structure understanding. Our model achieves significant gains of +6.2%, +5.7%, and +25.6%, respectively, over MMTab-trained baselines, while producing transparent and verifiable reasoning traces. These results establish CoReTab as a robust and generalizable supervision framework for improving multi-step reasoning in multimodal table understanding.

AI-text detectors achieve high accuracy on in-domain benchmarks, but often struggle to generalize across different generation conditions such as unseen prompts, model families, or domains. While prior work has reported these generalization gaps, there are limited insights about the underlying causes. In this work, we present a systematic study aimed at explaining generalization behavior through linguistic analysis. We construct a comprehensive benchmark that spans 6 prompting strategies, 7 large language models (LLMs), and 4 domain datasets, resulting in a diverse set of human- and AI-generated texts. Using this dataset, we fine-tune classification-based detectors on various generation settings and evaluate their cross-prompt, cross-model, and cross-dataset generalization. To explain the performance variance, we compute correlations between generalization accuracies and feature shifts of 80 linguistic features between training and test conditions. Our analysis reveals that generalization performance for specific detectors and evaluation conditions is significantly associated with linguistic features such as tense usage and pronoun frequency.

In this paper, we study the expansion of pluralia tantum, i.e., defective nouns which lack a singular form, like scissors. We base our work on an annotation framework specifically developed for the study of lexicalization of pluralia tantum, namely Lexicalization profiles. On a corresponding hand-annotated testset, we show that the OpenAI and DeepSeek models provide useful annotators for semantic, syntactic and sense categories, with accuracy ranging from 51% to 89%, averaged across all feature groups and languages. Next, we turn to a large-scale investigation of pluralia tantum. Using dictionaries, we extract candidate words for Italian, Russian and English and keep those for which the changing ratio of singular and plural form is evident in a corresponding reference corpus. We use an LLM to annotate each instance from the reference corpora according to the annotation framework. We show that the large amount of automatically annotated sentences for each feature can be used to perform in-depth linguistic analysis. Focusing on the correlation between an annotated feature and the grammatical form (singular vs. plural), patterns of morpho-semantic change are noted.

Integrating external knowledge into Large Language Models (LLMs) iscrucial for many real-world applications, yet current methods like Retrieval-Augmented Generation (RAG) face limitations with broad, multi-source queries, while long-context models are computationally prohibitive.We introduce CacheNotes: Task-Aware Key-Value Cache Compression. Given a task description and a corpus, CacheNotes first generates a sequence of Compression-Planning-Tokens (CPTs), an offline task-focused distillation pass that identifies and organizes key information from the corpus. These CPTs are then used to guide a one-time compression of the corpus into a compact, reusable KV cache, which is then used alone at inference time to efficiently answer diverse, reasoning-intensive queries, eliminating repeated retrieval or context expansion.Experiments on LongBench show that, on Question-Answering tasks at a 20× compression, CacheNotes outperforms RAG by over 8 F1 points and reduces latency by over 4×. On RULER, it surpasses previous query-agnostic compression methods by 55 points, narrowing the gap to query-aware compression approaches. Additional results on real-world enterprise and synthetic datasets demonstrate its strong performance on multi-hop and broad-coverage queries.

Large language models (LLMs) have shown strong capabilities as task-solving agents across interactive domains. However, in complex environments, these agents may need to rely on auxiliary guidance to reduce the search space or make up for limited domain-specific knowledge. Such guidance includes human-provided manuals and demonstrations, retrieved examples from memory or external tools, high-level heuristics, and agent-acquired knowledge from prior interactions. However, this guidance may be imperfect. For example, due to changes in the environment, ambiguous or simplified language, or retrieval errors from external sources, guidance can be incomplete, outdated, or contextually mismatched, potentially causing errors or failures during task execution. To address this, we introduce MIRAGE, a benchmark for MeasurIng Robustness of LLM Agents under Imperfect GuidancE. MIRAGE includes procedurally generated environments in navigation, cooking, and gaming, where both the environment and the auxiliary guidance vary in fidelity and relevance. We further extend MIRAGE to realistic web tasks via WebArena, using noisy or underspecified instructions extracted from demonstrations. Our findings reveal critical failure modes in current LLM agents and motivate future work on improving their robustness under imperfect guidance.

This paper presents a mechanistic investigation of how large language models (LLMs) generate contrastive sentiments. We define this task as transforming the sentiment of a given text (e.g., from positive to negative) while making minimal changes to its content. We identify two core mechanisms: (1) a preservation mechanism that maintains the sentiment of the input text, primarily mediated by specific attention heads, and (2) a sentiment transformation mechanism, which integrates a representation of the target sentiment label with the original valenced words using a circuit containing both MLP and attention layers. Building on these findings, we propose and validate a novel mechanistic intervention. By modifying key attention heads, we steer the LLM toward more effective contrastive generation, increasing the sentiment flip rate without sacrificing the minimality of changes. Our work not only deepens the understanding of the mechanisms underlying contrastive sentiment generation in LLMs, but also introduces a promising new direction to steer LLM behavior via targeted, mechanistic interventions.

In continual learning, our aim is to learn a new task without forgetting what was learned previously. In topic models, this translates to learning new topic models without forgetting previously learned topics. Previous work either considered Dynamic Topic Models (DTMs), which learn the evolution of topics based on the entire training corpus at once, or Online Topic Models, which are updated continuously based on new data but do not have long-term memory. To fill this gap, we propose the Continual Neural Topic Model (CoNTM), which continuously learns topic models at subsequent time steps without forgetting what was previously learned. This is achieved using a global prior distribution that is continuously updated. In our experiments, CoNTM consistently outperformed the dynamic topic model in terms of topic quality and predictive perplexity while being able to capture topic changes online. The analysis reveals that CoNTM can learn more diverse topics and better capture temporal changes than existing methods.

Recent advances in LLMs offer new opportunities for scalable, interactive mental health assessment, but excessive querying burdens users and is inefficient for real-world screening across transdiagnostic symptom profiles. We introduce MAQuA, a multi-outcome modeling and adaptive question-asking framework for simultaneous, multidimensional mental health screening. Combining multi-outcome modeling on language responses with item response theory (IRT) and factor analysis, MAQuA selects the questions with most informative responses across multiple dimensions at each turn to optimize diagnostic information, improving accuracy and potentially reducing response burden. Empirical results on a novel dataset reveal that MAQuA reduces the number of assessment questions required for score stabilization by 50–87% compared to random ordering (e.g., achieving stable depression scores with 71% fewer questions and eating disorder scores with 85% fewer questions). MAQuA demonstrates robust performance across both internalizing (depression, anxiety) and externalizing (substance use, eating disorder) domains, with early stopping strategies further reducing patient time and burden. These findings position MAQuA as a powerful and efficient tool for scalable, nuanced, and interactive mental health screening, advancing the integration of LLM-based agents into real-world clinical workflows.

Inspired by their success in image synthesis, diffusion models offer a flexible, iterative alternative to rigid left-to-right text generation. However, a fundamental training-inference discrepancy hinders their performance: models are trained on corrupted ground-truth tokens, but at inference time they must denoise inputs corrupted from their own predictions. To bridge this gap, we propose a unified framework. First, Deeper Self-Prediction (DSP) is a multi-step training objective that teaches robust self-correction by forcing the model to denoise its own intermediate outputs. Second, UCB-guided Decoding is a principled inference algorithm that frames token re-masking as a multi-armed bandit problem, using the Upper Confidence Bound (UCB) to balance exploration and exploitation. Experiments on text generation tasks demonstrate consistent improvements over existing diffusion baselines. The framework achieves higher faithfulness and coherence according to both automatic metrics and LLM-as-a-Judge evaluations.

Language identification (LID) is a critical step in curating multilingual LLM pretraining corpora from web crawls. While many studies on LID model training focus on collecting diverse training data to improve performance, low-resource languages – often limited to single-domain data, such as the Bible – continue to perform poorly. To resolve these class imbalance and bias issues, we propose a novel supervised contrastive learning (SCL) approach to learn domain-invariant representations for low-resource languages. We show that our approach improves LID performance on out-of-domain data for low-resource languages by 3.2 percentage points, while maintaining its performance for the high-resource languages.

Logical fallacies are common in public communication and can mislead audiences; fallacious arguments may still appear convincing despite lacking soundness, because convincingness is inherently subjective. We present the first computational study of how emotional framing interacts with fallacies and convincingness, using large language models (LLMs) to systematically change emotional appeals in fallacious arguments. We benchmark eight LLMs on injecting emotional appeal into fallacious arguments while preserving their logical structures, then use the best models to generate stimuli for a human study. Our results show that LLM-driven emotional framing reduces human fallacy detection in F1 by 14.5% on average. Humans perform better in fallacy detection when perceiving enjoyment than fear or sadness, and these three emotions also correlate with significantly higher convincingness compared to neutral or other emotion states. Our work has implications for AI-driven emotional manipulation in the context of fallacious argumentation.

Text-to-Visualization (Text2Vis) systems translate natural language queries over tabular data into concise answers and executable visualizations. While closed-source LLMs generate functional code, the resulting charts often lack semantic alignment and clarity—qualities that can only be assessed post-execution. Open-source models struggle even more, frequently producing non-executable or visually poor outputs. Although supervised fine-tuning can improve code executability, it fails to enhance overall visualization quality, as traditional SFT loss cannot capture post-execution feedback. To address this gap, we propose RL-Text2Vis, the first reinforcement learning framework for Text2Vis generation. Built on Group Relative Policy Optimization (GRPO), our method uses a novel multi-objective reward that jointly optimizes textual accuracy, code validity, and visualization quality using post-execution feedback. By training Qwen2.5 models (7B and 14B), RL-Text2Vis achieves a 22% relative improvement in chart quality over GPT-4o on the Text2Vis benchmark and boosts code execution success from 78% to 97% relative to its zero-shot baseline. Our models significantly outperform strong zero-shot and supervised baselines and also demonstrate robust generalization to out-of-domain datasets like VIS-Eval and NVBench. These results establish GRPO as an effective strategy for structured, multimodal reasoning in visualization generation. We release our code at https://github.com/vis-nlp/RL-Text2Vis.

Aligning Large Language Models (LLMs) with human preferences is crucial, but standard methods like Reinforcement Learning from Human Feedback (RLHF) are often complex and unstable. In this work, we propose a new, simpler approach that recasts alignment through the lens of Maximum Marginal Likelihood (MML) estimation. Our new MML-based Preference Optimization (MMPO) maximizes the marginal log-likelihood of a preferred text output, using the preference pair as samples for approximation, and forgoes the need for both an explicit reward model and entropy maximization. We theoretically demonstrate that MMPO implicitly performs preference optimization, producing a weighted gradient that naturally up-weights chosen responses over rejected ones. Across models ranging from 135M to 8B parameters, we empirically show that MMPO: 1) is more stable with respect to the hyperparameter compared to alternative baselines, and 2) achieves competitive or superior preference alignment while better preserving the base model’s general language capabilities. Through a series of ablation experiments, we show that this improved performance is indeed attributable to MMPO’s implicit preference optimization within the gradient updates.

We examine in how far a person’s value system has an impact on their perception of offensiveness. For instance, a scholar is likely to be offended by being accused of reporting unverified claims whereas many non-scholars would not feel that way. Thus, we move away from the assumption that offensiveness can be defined through a universal perspective. Ultimately, such research aims to support personalized approaches to content moderation. Our main contribution is the introduction of a dataset consisting of neutrally-phrased sentences on controversial topics, evaluated by individuals from 4 different value systems. This allows us to identify offensiveness patterns across value systems and conduct classification experiments.

Instruction tuning is a widely used approach to improve the instruction-following ability of large language models (LLMs). Instruction-tuning datasets typically include a mixture of context-augmented and context-free examples, yet prior work has largely combined these data types without examining their distinct effects. In this paper, we investigate how training LLMs with or without context affects model behavior and downstream performance. First, in the text domain, we show that LLMs trained with context attend more strongly to the provided knowledge, achieving better grounding. We also observe that context-augmented training shifts how LLMs use knowledge: models store and leverage less on parametric knowledge and instead depend more on the provided context. Second, we observe that using LLM trained with context-augmented data as the backbone for vision-language models reduces hallucination and improves grounding in the visual domain. Finally, we explore practical strategies for real-world deployments where context availability varies. We show that maintaining separate context-augmented and context-free models and routing inputs between them yields more robust overall performance than training a single mixed model, as it better preserves their complementary strengths.

The ability of language models in RAG systems to selectively refuse to answer based on flawed context is critical for safety, yet remains a significant failure point. Our large-scale study reveals that even frontier models struggle in this setting, with refusal accuracy dropping below 50% on multi-document tasks, while exhibiting dangerous over-confidence or over-caution. Static benchmarks fail to reliably evaluate this capability, as models exploit dataset-specific artifacts and memorize test instances. We introduce RefusalBench, a generative methodology that programmatically creates diagnostic test cases through controlled linguistic perturbation. Our framework employs 176 distinct perturbation strategies across six categories of informational uncertainty and three intensity levels. Evaluation of over 30 models uncovers systematic failure patterns: refusal comprises separable detection and categorization skills, and neither scale nor extended reasoning improves performance. We find that selective refusal is a trainable, alignment-sensitive capability, offering a clear path for improvement. We release two benchmarks—RefusalBench-NQ (single-document) and RefusalBench-GaRAGe (multi-document), and our complete generation framework to enable continued, dynamic evaluation of this critical capability.

Retrieval-augmented generation (RAG) systems address complex user requests by decomposing them into subqueries, retrieving potentially relevant documents for each, and then aggregating them to generate an answer. Efficiently selecting informative documents requires balancing a key trade-off: (i) retrieving broadly enough to capture all the relevant material, and (ii) limiting retrieval to avoid excessive noise and computational cost. We formulate query decomposition and document retrieval in an exploitation-exploration setting, where retrieving one document at a time builds a belief about the utility of a given sub-query and informs the decision to continue exploiting or exploring an alternative. We experiment with a variety of bandit learning methods and demonstrate their effectiveness in dynamically selecting the most informative sub-queries. Our main finding is that estimating document relevance using rank information and human judgments yields a 35% gain in document-level precision, 15% increase in α-nDCG, and better performance on the downstream task of long-form generation. Code is available on GitHub.

Vision-Language Models (VLMs) have shown strong multimodal reasoning capability on Visual-Question-Answering (VQA) benchmarks. However, their robustness against textual misinformation remains under-explored. While existing research has extensively studied the effect of misinformation in text-only domains, it is not clear how VLMs arbitrate between contradictory information from different modalities. To bridge the gap, we first propose the ConText-VQA (i.e. Conflicting Text) dataset, consisting of image-question pairs together with systematically generated persuasive prompts that deliberately conflict with visual evidence. Then, a thorough testing framework is designed and executed to benchmark the susceptibility of various models to these conflicting textual inputs. Comprehensive experiments over 11 state-of-the-art VLMs reveal that these models are indeed vulnerable to misleading prompts, often overriding clear visual evidence in favor of the conflicting text, and show an average performance drop of over 48.2% after only one round of persuasive conversation. Our findings highlight a critical limitation in current VLMs and underscore the need for improved robustness against textual manipulation.

We find that correct-to-incorrect sycophancy signals are most linearly accessible within multi-head attention activations. Motivated by the linear representation hypothesis, we train linear probes across the residual stream, multilayer perceptron (MLP), and attention layers to analyze where these signals emerge. Although separability appears in the residual stream and MLPs, steering using these probes is most effective in a sparse subset of middle-layer attention heads. Using TruthfulQA as the base dataset, we find that probes trained on it transfer effectively to other factual QA benchmarks. Furthermore, comparing our discovered direction to previously identified “truthful” directions reveals limited overlap, suggesting that factual accuracy, and deference resistance, arise from related but distinct mechanisms. Attention-pattern analysis further indicates that the influential heads attend disproportionately to expressions of user doubt, contributing to sycophantic shifts. Overall, these findings suggest that sycophancy can be mitigated through simple, targeted linear interventions that exploit the internal geometry of attention activations. Code will be released upon publication.

Large language models (LLMs) are increasingly capable of generating functional source code, raising concerns about authorship, accountability, and security. While detecting AI-generated code is critical, existing datasets and benchmarks are narrow, typically limited to binary human–machine classification under in-distribution settings. To bridge this gap, we introduce AICD Bench, the most comprehensive benchmark for AI-generated code detection. It spans 2M examples, 77 models across 11 families, and 9 programming languages, including recent reasoning models. Beyond scale, AICD Bench introduces three realistic detection tasks: (i) Robust Binary Classification under distribution shifts in language and domain, (ii) Model Family Attribution, grouping generators by architectural lineage, and (iii) Fine-Grained Human–Machine Classification across human, machine, hybrid, and adversarial code. Extensive evaluation on neural and classical detectors shows that performance remains far below practical usability, particularly under distribution shift and for hybrid or adversarial code. We release AICD Bench as a unified, challenging evaluation suite to drive the next generation of robust approaches for AI-generated code detection. The data and the code are available at https://huggingface.co/AICD-bench.

The potential misuse and misalignment of language models (LMs) is a central safety concern. This work presents Self-Destruct, a novelmechanism to restrict specific behaviors in LMs by leveraging overlooked properties of the underlying hardware. We observe that the LMframeworks use limited-precision formats (e.g., BF16), which are vulnerable to overflow errors during matrix multiplications. Exploitingthis property, Self-Destruct replaces selected weights in pre-trained LM layers with values that act as traps, triggering a system error onlywhen the model engages in targeted behaviors, such as harmful text generation, while leaving normal functionality unaffected. Unlike posthoc filters, this safeguard is embedded directly within the model, introduces neither inference overhead nor auxiliary models, and requires only a set of examples for calibration. Extensive experiments with five LM families demonstrate that Self-Destruct provides competitive protection against jailbreak attacks while preserving accuracy on standard benchmarks. In addition, we also show that Self-Destruct is versatile, helping mitigate biased text generation and enable model fingerprinting, highlighting the potential of hardware-aware safeguards as an efficient, low-overhead complement to existing LM defenses.

Large language models (LLMs) are known to "hallucinate" by generating false or misleading outputs. Hallucinations pose various harms, from erosion of trust to widespread misinformation. Existing hallucination evaluation, however, focuses only on correctness and often overlooks consistency, necessary to distinguish and address these harms. To bridge this gap, we introduce prompt multiplicity, a framework for quantifying consistency in LLM evaluations. Our analysis reveals significant multiplicity (over 50% inconsistency in benchmarks like Med-HALT), suggesting that hallucination-related harms have been severely misunderstood. Furthermore, we study the role of consistency in hallucination detection and mitigation. We find that: (a) detection techniques detect consistency, not correctness, and (b) mitigation techniques like RAG, while beneficial, can introduce additional inconsistencies. By integrating prompt multiplicity into hallucination evaluation, we provide an improved framework of potential harms and uncover critical limitations in current detection and mitigation strategies.

In science, promotional language (’hype’) is increasing and can undermine objective evaluation of evidence, impede research development, and erode trust in science. In this paper, we introduce the task of automatic detection of hype, which we define as hyperbolic or subjective language that authors use to glamorize, promote, embellish, or exaggerate aspects of their research. We propose formalized guidelines for identifying hype language and apply them to annotate a portion of the National Institutes of Health (NIH) grant application corpus. We then evaluate traditional text classifiers and language models on this task, comparing their performance with a human baseline. Our experiments show that formalizing annotation guidelines can help humans reliably annotate candidate hype adjectives and that using our annotated dataset to train machine learning models yields promising results. Our findings highlight the linguistic complexity of the task and the potential need for domain knowledge. While some linguistic works address hype detection, to the best of our knowledge, we are the first to approach it as a natural language processing task. Our annotation guidelines and dataset are available at https://github.com/hype-busters/eacl2026-hype-dataset.

The alignment of pre-trained LLMs continues to draw significant attention from both industry and academia, aiming to ensure responses that are helpful, harmless, and honest. However, identifying a point in the model’s representation subspace that simultaneously satisfies all these properties remains challenging. H3Fusion addresses this challenge by introducing a mixture-of-experts (MoE)-based fusion mechanism that models alignment as a controllable drift within the subspace, guided by a drift-regularization loss to balance competing alignment dimensions. Furthermore, we formulate the alignment by finding a dual objective of harnessing the distance of generated embeddings and alignment embeddings, and introduce gating loss by canalizing the activations on the contributing experts. Extensive evaluations of three benchmark datasets show that H3Fusion is more helpful, less harmful, and more honest in three aspects: it outperforms each individually aligned model by 11.37%, and provides stronger robustness compared to the state-of-the-art LLM ensemble approaches by 13.77% and model-merging approaches by 6.18 %. Code is available at https://github.com/git-disl/h3fusion.

We investigate how well large language models (LLMs) generalize across different task difficulties, a key question for effective data curation and evaluation. Existing research is mixed regarding whether training on easier or harder data leads to better results, and whether those gains come on easier or harder test data. We address this question by conducting a systematic evaluation of LLMs’ generalization across models, datasets, and fine-grained groups of example difficulty. We rank examples in six datasets using the outputs of thousands of different LLMs and Item Response Theory (IRT), a well-established difficulty metric in educational testing. Unlike prior work, our difficulty ratings are therefore determined solely by the abilities of many different LLMs, excluding human opinions of difficulty. With a more objective, larger-scale, and finer-grained analysis, we show that cross-difficulty generalization is often limited; training on either easy or hard data cannot achieve consistent improvements across the full range of difficulties. These results show the importance of having a range of difficulties in both training and evaluation data for LLMs, and that taking shortcuts with respect to difficulty is risky.

Enabling large language models (LLMs) to unlearn knowledge and capabilities acquired during training has proven vital for ensuring compliance with data regulations and promoting ethical practices in generative AI. Although there are growing interests in developing various unlearning algorithms, it remains unclear how to best formulate the unlearning problem. The most popular formulation uses a weighted sum of forget and retain loss, but it often leads to performance degradation due to the inherent trade-off between forget and retain losses. In this work, we argue that it is important to model the hierarchical structure of the unlearning problem, where the forget problem (which unlearns certain knowledge and/or capabilities) takes priority over the retain problem (which preserves model utility). This hierarchical structure naturally leads to a bi-level optimization formulation where the lower-level objective focuses on minimizing the forget loss, while the upper-level objective aims to maintain the model’s utility. Based on this new formulation, we propose a novel algorithm, termed Bi-Level UnleaRning (), which not only possesses strong theoretical guarantees but more importantly, delivers superior performance. In particular, our extensive experiments demonstrate that consistently outperforms all the state-of-the-art algorithms across various unlearning tasks, models, and metrics.

Hyperscaling of data and parameter count in LLMs is yielding diminishing improvement when weighed against training costs, underlining a growing need for more efficient finetuning and inference without sacrificing performance. This is especially so for multimodal language models (MLMs), where the overhead of processing multimodal tokens can limit their practical viability. Parallely, recent work has uncovered implicit cross-modal alignment in the deeper layers of large MLMs, deepening our understanding of how MLMs process and encode information. Motivated by this, and our observation that MLMs naturally defer most cross-modal token interactions to deeper layers of the model, we propose a simple modification. Instead of concatenation with the language prompt at the start, we insert multimodal tokens directly into the middle, allowing them to entirely bypass the early layers. Our results with diverse modalities, (i) LLaVA & BLIP for vision, (ii) LTU for audio, and (iii) MoLCA for molecular data, and model sizes, starting from 350M to 13B parameters, indicate that our method reduces both training and inference costs, while at least preserving, if not surpassing the performance of existing baselines.

Despite their impressive performance on complex tasks, current language models (LMs) typically operate in a vacuum: Each input query is processed separately, without retaining insights from previous attempts. Here, we present Dynamic Cheatsheet (DC), a lightweight framework that endows a black-box LM with a persistent, evolving memory. Rather than repeatedly re-discovering or re-committing the same solutions and mistakes, DC enables models to store and reuse accumulated strategies, code snippets, and general problem-solving insights at inference time. This test-time learning enhances performance substantially across a range of tasks without needing explicit ground-truth labels or human feedback. Leveraging DC, Claude 3.5 Sonnet’s accuracy more than doubled on AIME math exams once it began retaining algebraic insights across questions. Similarly, GPT-4o’s success rate on the Game of 24 puzzle increased from about 10% to 99% after the model discovered and reused a Python-based solution. In tasks prone to arithmetic mistakes, such as balancing equations, DC enabled GPT-4o and Claude to reach near-perfect accuracy by recalling previously validated code, whereas their baselines stagnated around 50%. Beyond arithmetic challenges, DC yields notable accuracy gains on knowledge-demanding tasks. Claude achieved a 9% improvement in GPQA-Diamond and an 8% boost on MMLU-Pro Engineering and Physics problems. Crucially, DC’s memory is self-curated, focusing on concise, transferable snippets rather than entire transcripts, thereby facilitating meta-learning and avoiding context ballooning. Unlike fine-tuning or static retrieval methods, DC adapts LMs’ problem-solving skills on the fly, without modifying their underlying parameters, and offers a practical approach for continuously refining responses and cutting routine errors. Overall, our findings present DC as a promising approach for augmenting LMs with persistent memory, bridging the divide between isolated inference events and the cumulative, experience-driven learning characteristic of human cognition.

Quantifying uncertainty in large language models (LLMs) is crucial for applications where safety is a concern, as it helps identify factually incorrect LLM answers, commonly referred to as hallucinations. Recently, advancements have been made in quantifying uncertainty, specifically by incorporating the semantics of sampled answers to estimate entropy. These methods typically rely on a normalized probability that is calculated using a limited number of sampled answers. However, we note these estimation methods fail to account for the effects of the semantics that are possible to be obtained as answers, but are not observed in the sample. This is a significant oversight, since a heavier tail of unobserved answer probabilities indicates a higher level of overall uncertainty. To alleviate this issue, we propose Evidential Semantic Entropy (EVSE), which leverages evidence theory to represent both total ignorance arising from unobserved answers and partial ignorance stemming from the semantic relationships among the observed answers. Experiments show that EVSE significantly improves uncertainty quantification performance. Our code is available at: https://github.com/lucieK-J/EvidentialSemanticEntropy.git.

Advances in large language models (LLMs) have enabled significant capabilities in audio processing, resulting in state-of-the-art models now known as Large Audio Language Models (LALMs). However, minimal work has been done to measure audio understanding beyond automatic speech recognition (ASR). This paper closes that gap by proposing a benchmark suite, SCENEBench (Spatial, Cross-lingual, Environmental, Non-speech Evaluation), that targets a broad form of audio comprehension across four real-world categories: background sound understanding, noise localization, cross-linguistic speech understanding, and vocal characterizer recognition. In addition to performance, we also measure model latency. The purpose of this benchmark suite is to assess the audio beyond just what words are said— rather, in how they are said and the non-speech components of the audio. To strengthen ecological validity, we include a small human-recorded evaluation split per category. Based on the needs articulated by audio understanding use-cases of accessibility technology and industrial noise monitoring, this benchmark reveals critical gaps in current LALMs. The performance in each task is quite varied, with some tasks having performance far below random chance and others with high accuracy. We also provide a structured error taxonomy to characterize standard failure modes across tasks. These results provide direction for targeted improvements in model capabilities.

Large language models (LLMs) have demonstrated impressive performance on reasoning-intensive tasks, but enhancing their reasoning abilities typically relies on either reinforcement learning (RL) with verifiable signals or supervised fine-tuning (SFT) with high-quality long chain-of-thought (CoT) demonstrations, both of which are expensive. In this paper, we study a novel problem of incentivizing the reasoning capacity of LLMs without expensive high-quality demonstrations and reinforcement learning. We investigate whether the reasoning capabilities of LLMs can be effectively incentivized via supervision from significantly weaker models. We further analyze when and why such weak supervision succeeds in eliciting reasoning abilities in stronger models. Our findings show that supervision from significantly weaker reasoners can substantially improve student reasoning performance, recovering close to 94% of the gains of expensive RL at a fraction of the cost. Experiments across diverse benchmarks and model architectures demonstrate that weak reasoners can effectively incentivize reasoning in stronger student models, consistently improving performance across a wide range of reasoning tasks. Our results suggest that this simple weak-to-strong paradigm is a promising and generalizable alternative to costly methods for incentivizing strong reasoning capabilities at inference-time in LLMs. Code is at https://github.com/W2SR-ARR/Code.

Merging fine-tuned models is a promising alternative to costly multi-task training, but task interference remains a challenge, especially as the number of tasks grows. We present DivMerge, a reference-free method that merges models trained on different tasks by minimizing Jensen-Shannon divergence between their outputs and those of the merged model, automatically balancing task importance. While the method exhibits strong theoretical properties, experiments on classification and generative tasks with autoregressive models show that DivMerge consistently outperforms prior work, and remains robust when scaling to more tasks.

Watermarking has emerged as a promising solution for tracing and authenticating text generated by large language models (LLMs). A common approach to LLM watermarking is to construct a green/red token list and assign higher or lower generation probabilities to the corresponding tokens, respectively. However, most existing watermarking algorithms rely on heuristic green/red token list designs, as directly optimizing the list design with techniques such as reinforcement learning (RL) comes with several challenges. First, desirable watermarking involves multiple criteria, i.e., detectability, text quality, robustness against removal attacks, and security against spoofing attacks. Directly optimizing for these criteria introduces many partially conflicting reward terms, leading to an unstable convergence process. Second, the vast action space of green/red token list choices is susceptible to reward hacking. In this paper, we propose an end-to-end RL framework for robust and secure LLM watermarking. Our approach adopts an anchoring mechanism for reward terms to ensure stable training and introduces additional regularization terms to prevent reward hacking. Experiments on standard benchmarks with two backbone LLMs show that our method achieves a state-of-the-art trade-off across all criteria, with notable improvements in resistance to spoofing attacks without degrading other criteria.

LLM agents are increasingly deployed to plan, retrieve, and write with tools, yet evaluation still leans on static benchmarks and small human studies. We present the Agent-Testing Agent (ATA), a meta-agent that combines static code analysis, developer interrogation, literature mining, and persona-driven adversarial test generation whose difficulty adapts via judge feedback. Each dialogue is scored with an LLM-as-a-Judge (LAAJ) rubric and used to steer subsequent tests toward the agent’s weakest capabilities. On a travel planner and a Wikipedia writer, the ATA surfaces more diverse and severe failures than expert annotators while matching severity, and finishes in 20–30 minutes versus ten-annotator rounds that took days. Ablating code analysis and web search increases variance and miscalibration, underscoring the value of evidence-grounded test generation. The ATA outputs quantitative metrics and qualitative bug reports for developers. We release the full open-source implementation.

Attribution and fact verification are critical challenges in natural language processing for assessing information reliability. While automated systems and Large Language Models (LLMs) aim to retrieve and select concise evidence to support or refute claims, they often present users with either insufficient or overly redundant information, leading to inefficient and error-prone verification. To address this, we propose Evidence Ranking, a novel task that prioritizes presenting sufficient information as early as possible in a ranked list. This minimizes user reading effort while still making all available evidence accessible for sequential verification. We compare two approaches for the new ranking task: one-shot ranking and incremental ranking. We introduce a new evaluation framework, inspired by information retrieval metrics, and construct a unified benchmark by aggregating existing fact verification datasets. Extensive experiments with diverse models show that incremental ranking strategies better capture complementary evidence and that LLM-based methods outperform shallower baselines, while still facing challenges in balancing sufficiency and redundancy. Compared to evidence selection, we conduct a controlled user study and demonstrate that evidence ranking both reduces reading effort and improves verification. This work provides a foundational step toward more interpretable, efficient, and user-aligned information verification systems.

We present a comprehensive evaluation of large language models’ (LLMs) ability to process culturally grounded language, specifically to understand and pragmatically use figurative expressions that encode local knowledge and social nuance. Using figurative language as a proxy for cultural nuance and local knowledge, we design evaluation tasks for contextual understanding, pragmatic use, and connotation interpretation across Arabic and English. We evaluate 22 open- and closed-source LLMs on Egyptian Arabic idioms, multidialectal Arabic proverbs, and English proverbs. Results show a consistent hierarchy: accuracy on Arabic proverbs is 4.29% lower than on English proverbs, and performance on Egyptian idioms is 10.28% lower than on Arabic proverbs. On the pragmatic use task, accuracy drops by 14.07% relative to understanding, though providing idioms’ contextual sentences improves accuracy by 10.66%. Models also struggle with connotative meaning, reaching at most 85.58% agreement with human annotators on idioms with full inter-annotator agreement. Figurative language thus serves as an effective diagnostic for cultural reasoning, revealing that while LLMs often interpret figurative meaning, they still face major challenges in using it appropriately. To support future research, we release Kinayat, the first dataset of Egyptian Arabic idioms designed for both figurative understanding and pragmatic use evaluation.

Machine translation (MT) systems universally degrade when faced with code-mixed text. This problem is more acute for low-resource languages that lack dedicated parallel corpora. This work directly addresses this gap for Vietnamese-English, a language context characterized by challenges including orthographic ambiguity and the frequent omission of diacritics in informal text. We introduce VietMix, the first expert-translated, naturally occurring parallel corpus of Vietnamese-English code-mixed text. We establish VietMix’s utility by developing a data augmentation pipeline that leverages iterative fine-tuning and targeted filtering. Experiments show that models augmented with our data outperform strong back-translation baselines by up to +3.5 xCOMET points and improve zero-shot models by up to +11.9 points. Our work delivers a foundational resource for a challenging language pair and provides a validated, transferable framework for building and augmenting corpora in other low-resource settings.

Multimodal Large Language Models (MLLMs) show reasoning promise, yet their visual perception is a critical bottleneck. Paradoxically, MLLMs sometimes produce correct answers while misinterpreting crucial visual elements, masking these underlying perception failures. Our preliminary analysis on a joint perception-reasoning dataset revealed that 29% of correct reasoning answers from a leading MLLM contained perception errors. To systematically study visual perception abilities of MLLMs, we introduce Do You See Me- a scalable, programmatically generated benchmark with 1758 images and 2612 questions across seven core subtasks spanning 2D and 3D variants (twelve total tasks) providing parametric control over difficulty levels. The benchmark tasks are inspired by human psychology. Our evaluation of eleven leading MLLMs reveals a stark deficit: humans achieve 95.83% accuracy, while top MLLMs average below 50%. This performance gap widens drastically as task complexity increases. Further diagnostics show: (1) supervised finetuning offers only modest gains (11%), (2) models tend to exploit task “shortcuts” like MCQ formats over detailed visual analysis, and (3) Chain-of-Thought prompting can degrade complex visual tasks by verbalizing images into lossy text. These findings expose the foundational perception limits in current MLLMs and highlight the need for robust visual perception improvements in MLLMs. The benchmark dataset, source code and evaluation scripts are available at[<https://github.com/microsoft/Do-You-See-Me>].

Large Language Models (LLMs) increasingly mediate our social, cultural, and political interactions. While they can simulate some aspects of human behavior and decision-making, it is still underexplored whether repeated interactions with other agents amplify their biases or lead to exclusionary behaviors. To this end, we study Chirper.ai—an LLM-driven social media platform—analyzing 7M posts and interactions among 32K LLM agents over a year. We start with homophily and social influence among LLMs, learning that similar to humans’, their social networks exhibit these fundamental phenomena. Next, we study the toxic language of LLMs, its linguistic features, and their interaction patterns, finding that LLMs show different structural patterns in toxic posting than humans. After studying the ideological leaning in LLMs posts, and the polarization in their community, we focus on how to prevent their potential harmful activities. We present a simple yet effective method, called Chain of Social Thought (CoST), that reminds LLM agents to avoid harmful posting.

Large Language Models (LLMs) are increasingly embedded in the daily lives of individuals across diverse social classes. This widespread integration raises urgent concerns about the subtle, implicit biases these models may contain. In this work, we investigate such biases through the lens of ethical reasoning, analyzing model responses to scenarios in a new dataset we propose comprising 1,016 scenarios, systematically categorized into ethical, unethical, and neutral types. Our study focuses on dimensions that are socially influential but less explored, including (i) residency status, (ii) political ideology, (iii) Fitness Status, (iv) educational attainment, and (v) attitudes toward AI. To assess LLMs’ behavior, we propose a baseline and employ one statistical test and one metric: a permutation test that reveals the presence of bias by comparing the probability distributions of ethical/unethical scenarios with the probability distribution of neutral scenarios on each demographic group, and a tendency measurement that captures the magnitude of bias with respect to the relative difference between probability distribution of ethical and unethical scenarios. Our evaluations of 12 prominent LLMs reveal persistent and nuanced biases across all four attributes, and Llama models exhibited the most pronounced biases. These findings highlight the need for refined ethical benchmarks and bias-mitigation tools in LLMs.

A wide array of NLP/NLU models have been developed for the Persian language and have shown promising results. However, the performance of such models drops significantly when applied to the colloquial form of Persian. This challenge arises from the substantial differences between colloquial and formal Persian and the lack of parallel data facilitating the robustness of the model to the colloquial data or to transform the data to formal Persian. In addressing this gap, our research is dedicated to the development of the HarfoSokhan dataset, a large-scale colloquial to formal Persian parallel dataset of 6M sentence pairs. Our proposed dataset is a critical resource for training models that can effectively bridge the linguistic variations between colloquial and formal Persian. To illustrate the utility of our dataset, we used it to train a GPT2 model, which exhibited remarkable proficiency in colloquial to formal text style transfer, outperforming both OpenAI’s GPT-3.5-turbo model and a leading rule-based system in this task. This conclusion is supported by our proposed ranking-based human evaluation. The results underscore the significance of the HarfoSokhan dataset in enhancing the performance of natural language processing models in the challenging task of colloquial to formal Persian conversion.

Language models (LMs) excel at tasks across diverse domains, yet require substantial computational resources during inference. Compression techniques such as pruning and quantization offer a practical path towards efficient LM deployment, exemplified by their ability to preserve performance on general-purpose benchmarks. However, general-purpose LM compression methods can negatively affect performance in specialized domains (e.g. biomedical or legal). Recent work has sought to address this issue, but requires a computationally expensive full-parameter fine-tuning pipeline. To this end, we propose MixCal, a novel calibration method designed to improve the in-domain performance of compressed LMs in a post-training setting. Through extensive experimentation, we demonstrate that MixCal substantially outperforms existing approaches on domain-specific tasks while preserving general performance. Notably, these performance gains are achieved while also reducing the computational cost of LM compression.

Personalization in LLMs often relies on costly human feedback or interaction logs, limiting scalability and neglecting deeper user attributes. We introduce (Generative Response with Aligned Values, Interests, and Traits of You), a framework for generating synthetic, profile-grounded preference data that captures users’ interests, values, beliefs, and personality traits. By integrating demographic, cultural, and psychological frameworks—including Hofstede’s cultural dimensions, Schwartz’s basic values, the World Values Survey, and Big Five OCEAN traits—synthesizes chosen/rejected preference pairs to guide personalized content generation. We evaluate on book descriptions for 400 Amazon users, comparing it to prompt-based conditioning, standard fine-tuning, and naive synthetic pair generation. Profile-grounded synthetic data consistently improves generation, especially across multiple cultures (USA, Brazil, Japan, India), achieving over 4% higher preference gains across baselines, with user studies showing that outputs are preferred over 86% of the time. Our results show that scenario-grounded synthetic data can capture richer user variation, reduce reliance on costly annotation, and produce more engaging, user-centered content, offering a scalable path for LLM personalization. Code and datasets will be released upon acceptance.

While multimodal large language models (LLMs) excel at dialogue, whether they can adequately parse the structure of conversation—conversational roles and threading—remains underexplored. In this work, we introduce a suite of tasks and release TV-MMPC, a new annotated dataset, for multimodal conversation structure understanding. Our evaluation reveals that while all multimodal LLMs outperform our heuristic baseline, even the best-performing model we consider experiences a substantial drop in performance when character identities of the conversation are anonymized. Beyond evaluation, we carry out a sociolinguistic analysis of 350,842 utterances in TVQA. We find that while female characters initiate conversations at rates in proportion to their speaking time, they are 1.2 times more likely than men to be cast as an addressee or side-participant, and the presence of side-participants shifts the conversational register from personal to social.

Psychological insights have long shaped pivotal NLP breakthroughs, from attention mechanisms to reinforcement learning and social modeling. As Large Language Models (LLMs) develop, there is a rising consensus that psychology is essential for capturing human-like cognition, behavior, and interaction.This paper reviews how psychological theories can inform and enhance stages of LLM development. Our review integrates insights from six subfields of psychology, including cognitive, developmental, behavioral, social, personality psychology, and psycholinguistics. With stage-wise analysis, we highlight current trends and gaps in how psychological theories are applied. By examining both cross-domain connections and points of tension, we aim to bridge disciplinary divides and promote more thoughtful integration of psychology into NLP research.

Large language model (LLM) routing has emerged as a crucial strategy for balancing computational costs with performance by dynamically assigning queries to the most appropriate model based on query complexity. Despite recent advances showing that preference-data-based routers can outperform traditional methods, current evaluation benchmarks remain limited—they largely focus on general model capabilities while overlooking task-specific behaviors and critical concerns such as privacy, safety, and potential backdoor vulnerabilities introduced through preference data. In response, we propose the DSC benchmark: Diverse, simple, and categorized, an evaluation framework that categorizes router performance across a broad spectrum of query types—including coding, translation, mathematics, human instructions, general knowledge, and LLM jailbreaking—and integrates privacy and safety assessments to reveal hidden risks. Our experiments on three preference-based routers and two commercial counterparts demonstrate that while these systems improve efficiency, they often make suboptimal, category-driven decisions; for instance, a BERT-based router directs all coding and mathematics queries to the most powerful LLM—even when simpler models would suffice—while routing jailbreaking attempts to weaker models, thereby elevating safety risks.

Diet plays a central role in human health, and Nutrition Question Answering (QA) offers a promising path toward personalized dietary guidance and the prevention of diet-related chronic diseases. However, existing methods face two fundamental challenges: the limited reasoning capacity of single-agent systems and the complexity of designing effective multi-agent architectures, as well as contextual overload that hinders accurate decision-making. We introduce Nutritional-Graph Router (NG-Router), a novel framework that formulates nutritional QA as a supervised, knowledge-graph–guided multi-agent collaboration problem. NG-Router integrates agent nodes into heterogeneous knowledge graphs and employs a graph neural network to learn task-aware routing distributions over agents, leveraging soft supervision derived from empirical agent performance. To further address contextual overload, we propose a gradient-based subgraph retrieval mechanism that identifies salient evidence during training, thereby enhancing multi-hop and relational reasoning. Extensive experiments across multiple benchmarks and backbone models demonstrate that NG-Router consistently outperforms both single-agent and ensemble baselines, offering a principled approach to domain-aware multi-agent reasoning for complex nutritional health tasks.

Large language model (LLM) agents are increasingly deployed to tackle complex tasks, often necessitating collaboration among multiple specialized agents. However, multi-agent collaboration introduces new challenges in planning, coordination, and verification. Execution failures frequently arise not from flawed reasoning alone, but from subtle misalignments in task interpretation, output format, or inter-agent handoffs. To address these challenges, we present VeriMAP, a framework for multi-agent collaboration with verification-aware planning. The VeriMAP planner decomposes tasks, models subtask dependencies, and encodes planner-defined passing criteria as subtask verification functions (VFs) in Python and natural language. We evaluate VeriMAP on diverse datasets, demonstrating that it outperforms both single- and multi-agent baselines while enhancing system robustness and interpretability. Our analysis highlights how verification-aware planning enables reliable coordination and iterative refinement in multi-agent systems, without relying on external labels or annotations.

Entity structure extraction, which aims to extract entities and their associated attribute–value structures from text, is an essential task for text understanding and knowledge graph construction. Existing methods based on large language models (LLMs) typically rely heavily on predefined entity attribute schemas or annotated datasets, often leading to incomplete extraction results. To address these challenges, we introduce ZOES, a novel approach to entity structure extraction that does not require any schema or annotated samples. ZOES operates via a principled mechanism of enrichment, refinement, and unification, based on the insight that an entity and its associated structure are mutually reinforcing. Experiments demonstrate that ZOES consistently enhances LLMs’ ability to extract more complete entity structures across three different domains, showcasing both the effectiveness and generalizability of the method. These findings suggest that such an enrichment, refinement, and unification mechanism may serve as a principled approach to improving the quality of LLM-based entity structure discovery in various scenarios.

In-context learning depends not only on what appears in the prompt but also on when it appears. To isolate this temporal component from semantic confounds, we construct prompts with repeated anchor tokens and average the model’s predictions over hundreds of random permutations of the intervening context. This approach ensures that any observed position-dependent effects are driven purely by temporal structure rather than token identity or local semantics. Across four transformer LLMs and three state-space/recurrent models, we observe a robust serial recall signature: models allocate disproportionate probability mass to the tokens that previously followed the anchor, but the strength of this signal is modulated by serial position, yielding model-specific primacy/recency profiles. We then introduce an overlapping-episode probe in which only a short cue from one episode is re-presented; retrieval is reliably weakest for episodes embedded in the middle of the prompt, consistent with "lost-in-the-middle" behavior. Mechanistically, ablating high-induction-score attention heads in transformers reduces serial recall and episodic separation. For state-space models, ablating a small fraction of high-attribution channels produces analogous degradations, suggesting a sparse subspace supporting induction-style copying. Together, these results clarify how temporal biases shape retrieval across architectures and provide controlled probes for studying long-context behavior.

Large-scale Vision-Language Models (LVLMs) are being deployed in real-world settings that require visual inference. As capabilities improve, applications in navigation, education, and accessibility are becoming practical. These settings require accommodation of perceptual variation rather than assuming a uniform visual experience. Color perception illustrates this requirement: it is central to visual understanding yet varies across individuals due to Color Vision Deficiencies, an aspect largely ignored in multimodal AI.In this work, we examine whether LVLMs can account for variation in color perception using the Ishihara Test. We evaluate model behavior through generation, confidence, and internal representation, using Ishihara plates as controlled stimuli that expose perceptual differences. Although models possess factual knowledge about color vision deficiencies and can describe the test, they fail to reproduce the perceptual outcomes experienced by affected individuals and instead default to normative color perception. These results indicate that current systems lack mechanisms for representing alternative perceptual experiences, raising concerns for accessibility and inclusive deployment in multimodal settings.

Extending Large Language Models (LLMs) to new languages is challenging, with most methods proposed suffering from high computational cost and catastrophic forgetting of original model capabilities. Embedding relearning (CITATION), a technique that creates new tokenizers and tunes embeddings on fixed model weights for target language adaptation, is both light-weight and performant. However, it has only been shown to work for older generation encoder-only models and for high resource languages. In this paper, we extend this framework to decoder-only LLMs focusing on joint adaptation to many languages, including low-resource ones. We experiment in three language groups over 100 languages each. We adapt a pre-trained LLM via switching to a customized tokenizer, and relearning the embedding layer. Across 96 diverse languages spanning both classification and generation tasks, we show embedding relearning improves models by up to 20%, being highly competitive with full-weight updating baselines while vastly more computationally efficient and mitigating catastrophic forgetting. This translates into better results in transferring the improved multilingual performance to tasks that build on core English abilities (e.g., multilingual math reasoning), compared to various baselines. Further analysis reveals the critical role of customizing tokenizers in achieving effective language transfer, particularly for non-Latin script languages.

Generalized Category Discovery (GCD) is a practical and challenging open-world task that aims to recognize both known and novel categories in unlabeled data using limited labeled data from known categories. Due to the lack of supervision, previous GCD methods face significant challenges, such as difficulty in rectifying errors for confusing instances, and inability to effectively uncover and leverage the semantic meanings of discovered clusters. Therefore, additional annotations are usually required for real-world applicability. However, human annotation is extremely costly and inefficient. To address these issues, we propose DeLFGCD , a unified framework for generalized category discovery that actively learns from diverse and collaborative LLM feedback. Our approach leverages three different types of LLM feedback to: (1) improve instance-level contrastive features, (2) generate category descriptions, and (3) align uncertain instances with LLM-selected category descriptions. Extensive experiments demonstrate the superior performance of DeLFGCD over state-of-the-art models across diverse datasets, metrics, and supervision settings.

The advent of complex, interconnected long-horizon LLM systems has made it incredibly tricky to identify where and when these systems break down. Evaluation capabilities that currently exist today are limited in that they often focus on simple metrics, end-to-end outcomes, and are dependent on the perspectives of humans. In order to match the increasing complexity of these many component systems, evaluation frameworks must also be able to reason, probe, iterate, and understand the nuanced logic passing through these systems. In this paper, we present RAFFLES, an offline evaluation architecture that incorporates iterative reasoning. Specifically, RAFFLES operates as an iterative, multi-component pipeline, using a central Judge to systematically identify faults and a set of specialized Evaluators to assess the quality of the candidate faults as well as rationales of the Judge. We evaluated RAFFLES with several benchmarks - the Who&When dataset to identify step-level faults in multi-agent systems and the ReasonEval datasets to diagnose step-level mathematical reasoning errors. RAFFLES outperforms strong baselines, achieving an accuracy of over 20% and 50% on the Who&When Hand-Crafted and Algorithmically-Generated datasets, and over 80% on the ReasonEval datasets. These results demonstrate a key step towards introducing automated fault detection for autonomous systems over labor-intensive manual review.

Large language models (LLMs) are safety-aligned to prevent harmful response generation, yet still remain vulnerable to jailbreak attacks. While prior works have focused on improving jailbreak attack effectiveness, they offer little explanation for why safety alignment fails. We address this gap by framing jailbreaks as inference-time alignment, connecting attack design and safety alignment within a unified optimization framework. This framing allows us to extend best-of-N inference-time alignment to the adversarial setting, called LIAR (Leveraging Inference-time Alignment to jailbReak), and derive suboptimality bounds that show LIAR provably approaches an optimal jailbreak as compute scales. Interestingly, our framework allows us to develop the notion of a Safety-Net, a measure of how vulnerable an LLM is to jailbreaks, which helps to explain why safety alignment can fail. Empirically, LIAR produces natural, hard-to-detect prompts that achieve a competitive attack success rate while running 10 to 100x faster than prior suffix-based jailbreaks.

Search-augmented large language models (LLMs) excel at knowledge-intensive tasks by integrating external retrieval. However, they often over-search – unnecessarily invoking search tool even when it does not improve response quality, which leads to computational inefficiency and hallucinations by incorporating irrelevant context. In this work, we conduct a systematic evaluation of over-searching across multiple dimensions, including query types, model categories, retrieval conditions, and multi-turn conversations. Our findings show: (i) search generally improves answer accuracy on answerable queries but harms abstention on unanswerable ones; (ii) over-searching is more pronounced in complex reasoning models and deep research systems, is exacerbated by noisy retrieval, and compounds across turns in multi-turn conversations; and (iii) the composition of retrieved evidence is crucial, as the presence of negative evidence improves abstention. To quantify over-searching, we introduce Tokens Per Correctness (TPC), an evaluation metric that captures the performance-cost trade-off for search-augmented LLMs. Lastly, we investigate mitigation approaches at both the query and retrieval levels and release the OverSearchQA benchmark to foster continued research into efficient search-augmented LLMs.

Evaluating creative writing generated by large language models (LLMs) remains challenging because open-ended narratives lack ground truths. Without performant automated evaluation methods, off-the-shelf (OTS) language models are employed as zero-shot judges, yet their reliability is unclear in this context. To address this gap, we introduce LitBench, a large-scale benchmark for creative writing evaluation, featuring a training corpus of 43,827 story pairs and a 2,480-pair test set curated from Reddit. Using LitBench, we benchmark existing LLM judges and train specialized reward models. Our analysis reveals that the strongest OTS judge, Claude-3.7-Sonnet, achieves only 73% agreement with human preferences. In contrast, our trained Bradley-Terry and generative reward models both reach 78% accuracy, outperforming all OTS judges. An online human study further validates our models, showing their rankings of newly generated stories align more closely with human preferences. Our work provides the first reliable benchmark and specialized reward models for creative writing, establishing a crucial foundation for the future development of more capable verifiers.

Long-context conversational agents require robust memory, but existing frameworks struggle to organize information effectively across dimensions like time and topic, leading to poor retrieval. To address this, we introduce H-Mem, a novel Hybrid Multi-Dimensional Memory architecture. H-Mem stores conversational facts in two parallel, hierarchical data structures: a temporal tree that organizes information chronologically and a semantic tree that organizes it conceptually. This dual-tree design enables a hybrid retrieval mechanism managed by an intelligent Mode Controller. Based on the query, the controller dynamically chooses between a sequential search using semantic anchors and an intersective search combining both hierarchies. Our experiments on long-context QA datasets demonstrate that H-Mem provides a more flexible approach to memory management, leading to significant improvements of over 8.4% compared to other state-of-the-art systems.

Group bias in Large Language Models (LLMs) is a well-documented issue, its impact in high-stakes domains such as personalized nutritional advice remains under explored. This study introduces the USChainMains dataset to systematically evaluate LLMs, prompting them with names associated with specific racial and gender groups and rigorously quantifying the healthfulness of the generated meal recommendations against established dietary standards. The findings demonstrate that LLMs systematically recommend meals with significantly higher levels of adverse nutrients for names associated with Black, Hispanic, or male individuals, thereby reflecting and potentially reinforcing detrimental dietary stereotypes. Furthermore, our analysis of two common mitigation strategies reveals their limitations. While model scaling improves overall recommendation healthfulness, it is insufficient to eliminate the healthfulness gap between demographic groups. Notably, while augmented reasoning was effective in mitigating gender bias, it did not mitigate racial disparities. This work underscores the necessity of developing more nuanced, group-aware debiasing techniques to ensure AI-driven systems advance, rather than hinder, health equity.

Transliteration has emerged as a promising means to bridge the gap between various languages in multilingual NLP, showing promising results especially for languages using non-Latin scripts. We investigate the degree to which shared script, overlapping token vocabularies, and shared phonology contribute to performance of multilingual models. To this end, we conduct controlled experiments using three kinds of transliteration (romanization, phonemic transcription, and substitution ciphers) as well as orthography. We evaluate each model on three downstream tasks—named entity recognition (NER), part-of-speech tagging (POS) and natural language inference (NLI)—and find that romanization significantly outperforms other input types in 7 out of 8 evaluation settings, largely consistent with our hypothesis that it is the most effective approach. We further analyze how each factor contributed to the success, and suggest that having longer (subword) tokens shared with pre-trained languages leads to better utilization of the model.