Large language models (LLM) have shown remarkable abilities in text generation, question answering, language translation, reasoning and many other tasks. It continues to advance rapidly and is becoming increasingly influential in various fields, from technology and business to education and entertainment. Despite LLM’s success in multiple areas, its ability to play abstract games, such as chess, is underexplored. Chess-playing requires the language models to output legal and reasonable moves from textual inputs. Here, we propose the Large language model ChessLLM to play full chess games. We transform the game into a textual format with the best move represented in the Forsyth-Edwards Notation. We show that by simply supervised fine-tuning, our model has achieved a professional-level Elo rating of 1788 in matches against the standard Elo-rated Stockfish when permitted to sample 10 times. We further show that data quality is important. Long-round data supervision enjoys a 350 Elo rating improvement over short-round data.

Dialect adapters that improve the performance of LLMs for NLU tasks on certain sociolects/dialects/national varieties (‘dialects’ for the sake of brevity) have been reported for encoder models. In this paper, we extend the idea of dialect adapters to decoder models in our architecture called LoRDD. Using MD-3, a publicly available dataset of word game-playing conversations between dialectal speakers, our task is Target Word Prediction (TWP) from a masked conversation. LoRDD combines task adapters and dialect adapters where the latter employ contrastive learning on pseudo-parallel conversations from MD-3. Our experiments on Indian English and Nigerian English conversations with two models (Mistral and Gemma) demonstrate that LoRDD outperforms four baselines on TWP. Additionally, it significantly reduces the performance gap with American English, narrowing it to 12% and 5.8% for word similarity, and 25% and 4.5% for accuracy, respectively. The focused contribution of LoRDD is in its promise for dialect adaptation of decoder models using TWP, a simplified version of the commonly used next-word prediction task.

The rise of LLMs has deflected a growing portion of human-computer interactions towards LLM-based chatbots.The remarkable abilities of these models allow users to interact using long, diverse natural language text covering a wide range of topics and styles. Phrasing these messages is a time and effort consuming task, calling for an autocomplete solution to assist users. We present **ChaI-TeA**: **Cha**t **I**n**te**raction **A**utocomplete; An autocomplete evaluation framework for LLM-based chatbot interactions. The framework includes a formal definition of the task, curated datasets and suitable metrics. We use it to evaluate 11 models on this task, finding that while current off-the-shelf models perform fairly, there is still much room for improvement, mainly in ranking of the generated suggestions. We provide insights for practitioners working on this task and open new research directions for researchers in the field. We release our framework to serve as a foundation for future research.

There has been increasing interest in building multilingual foundation models for NLP and speech research. This paper examines how to expand the speech translation capability of these models with restricted data. Whisper, a speech foundation model with strong performance on speech recognition and English translation, is used as the example model. Using speech-to-speech retrieval to analyse the audio representations generated by the encoder, we show that utterances from different languages are mapped to a shared semantic space. This shared embedding space can then be leveraged for zero-shot cross-lingual transfer in speech translation. By fine-tuning the Whisper decoder with only English-to-Chinese speech translation data, improved performance for translation to Chinese can be obtained for multiple languages, in addition to English. Furthermore, for languages related to those seen in training it is possible to perform speech translation, despite the model never seeing the language in training, or being able to perform transcription.

Question answering (QA)—giving correct answers to questions—is a popular task, but we test **reverse question answering (RQA)**: for an input answer, give a question with that answer. Past work tests QA and RQA separately, but we test them jointly, comparing their difficulty, aiding benchmark design, and checking reasoning consistency. We run 16 LLMs on QA and RQA with trivia questions/answers, revealing: 1) Versus RQA, LLMs are much less accurate in RQA for numerical answers, but slightly more accurate in RQA for textual answers; 2) LLMs often answer their own invalid questions from RQA accurately in QA, so RQA errors are not just from knowledge gaps; 3) RQA errors correlate with question difficulty and inversely correlate with answer frequencies in the Dolma corpus; and 4) LLMs struggle to give valid multi-hop questions. By finding question and answer types that lead to RQA errors, we suggest improvements for LLM reasoning.

AIs can beat humans in game environments; however, how helpful those agents are to human remains understudied. We augment Cicero, a natural language agent that demonstrates superhuman performance in Diplomacy, to generate both move and message advice based on player intentions. A dozen Diplomacy games with novice and experienced players, with varying advice settings, show that some of the generated advice is beneficial. It helps novices compete with experienced players and in some instances even surpass them. The mere presence of advice can be advantageous, even if players do not follow it.

In recent years, the use of prompts to guide the output of Large Language Models have increased dramatically. However, even the best of experts struggle to choose the correct words to stitch up a prompt for the desired task. To solve this, LLM driven prompt optimization emerged as an important problem. Existing prompt optimization methods optimize a prompt globally, where in all the prompt tokens have to be optimized over a large vocabulary while solving a complex task. The large optimization space (tokens) leads to insufficient guidance for a better prompt. In this work, we introduce Local Prompt Optimization (LPO) that integrates with any general automatic prompt engineering method. We identify the optimization tokens in a prompt and nudge the LLM to focus only on those tokens in its optimization step. We observe remarkable performance improvements on Math Reasoning (GSM8k and MultiArith) and BIG-bench Hard benchmarks across various automatic prompt engineering methods. Further, we show that LPO converges to the optimal prompt faster than global methods.

Reinforcement learning with human feedback (RLHF) is shown to largely benefit from precise reward models (RMs). However, recent studies in reward modeling schemes are skewed towards English, limiting the applicability of RLHF in multilingual alignments. In this work, we investigate the cross-lingual transfer of RMs trained in diverse languages, primarily from English. Our experimental results demonstrate the strong cross-lingual transfer of English RMs, exceeding target language RMs by 3~4% average increase in Multilingual RewardBench. Furthermore, we analyze the cross-lingual transfer of RMs through the representation shifts. Finally, we perform multilingual alignment to exemplify how cross-lingual transfer in RM propagates to enhanced multilingual instruction-following capability.

We propose selective debiasing – an inference-time safety mechanism designed to enhance the overall model quality in terms of prediction performance and fairness, especially in scenarios where retraining the model is impractical. The method draws inspiration from selective classification, where at inference time, predictions with low quality, as indicated by their uncertainty scores, are discarded. In our approach, we identify the potentially biased model predictions and, instead of discarding them, we remove bias from these predictions using LEACE – a post-processing debiasing method. To select problematic predictions, we propose a bias quantification approach based on KL divergence, which achieves better results than standard uncertainty quantification methods. Experiments on text classification datasets with encoder-based classification models demonstrate that selective debiasing helps to reduce the performance gap between post-processing methods and debiasing techniques from the at-training and pre-processing categories.

Current Large Language Models (LLMs) benchmarks are often based on open-ended or close-ended QA evaluations, avoiding the requirement of human labor. Close-ended measurements evaluate the factuality of responses but lack expressiveness. Open-ended capture the model’s capacity to produce discourse responses but are harder to assess for correctness. These two approaches are commonly used, either independently or together, though their relationship remains poorly understood. This work is focused on the healthcare domain, where both factuality and discourse matter greatly. It introduces a comprehensive, multi-axis suite for healthcare LLM evaluation, exploring correlations between open and close benchmarks and metrics. Findings include blind spots and overlaps in current methodologies. As an updated sanity check, we release a new medical benchmark–CareQA–, with both open and closed variants. Finally, we propose a novel metric for open-ended evaluations –Relaxed Perplexity– to mitigate the identified limitations.

Countless decisions shape our lives, and it is crucial to understand the how and why behind them. In this paper, we introduce a new LLM decision-making framework called STRUX, which enhances LLM decision-making by providing structured explanations. These include favorable and adverse facts related to the decision, along with their respective strengths. STRUX begins by distilling lengthy information into a concise table of key facts. It then employs a series of self-reflection steps to determine which of these facts are pivotal, categorizing them as either favorable or adverse in relation to a specific decision. Lastly, we fine-tune an LLM to identify and prioritize these key facts to optimize decision-making. STRUX has been evaluated on the challenging task of forecasting stock investment decisions based on earnings call transcripts and demonstrated superior performance against strong baselines. It enhances decision transparency by allowing users to understand the impact of different factors, representing a meaningful step towards practical decision-making with LLMs.

Document retrieval plays a crucial role in numerous question-answering systems, yet research has concentrated on the general knowledge domain and resource-rich languages like English. In contrast, it remains largely underexplored in low-resource languages and cross-lingual scenarios within specialized domain knowledge such as legal. We present a novel dataset designed for cross-lingual retrieval between Vietnamese and English, which not only covers the general domain but also extends to the legal field. Additionally, we propose auxiliary loss function and symmetrical training strategy that significantly enhance the performance of state-of-the-art models on these retrieval tasks. Our contributions offer a significant resource and methodology aimed at improving cross-lingual retrieval in both legal and general QA settings, facilitating further advancements in document retrieval research across multiple languages and a broader spectrum of specialized domains. All the resources related to our work can be accessed at huggingface.co/datasets/bkai-foundation-models/crosslingual.

Chiasmus, a debated literary device in Biblical texts, has captivated mystics while sparking ongoing scholarly discussion. In this paper, we introduce the first computational approach to systematically detect chiasmus within Biblical passages. Our method leverages neural embeddings to capture lexical and semantic patterns associated with chiasmus, applied at multiple levels of textual granularity (half-verses, verses). We also involve expert annotators to review a subset of the detected patterns. Despite its computational efficiency, our method achieves robust results, with high inter-annotator agreement and system accuracy of 0.80 at the verse level and 0.60 at the half-verse level. We further provide a qualitative analysis of the distribution of detected chiasmi, along with selected examples that highlight the effectiveness of our approach.

Rhetorical devices are difficult to translate, but they are crucial to the translation of literary documents. We investigate the use of multilingual embedding spaces to characterize the preservation of intertextuality, one common rhetorical device, across human and machine translation. To do so, we use Biblical texts, which are both full of intertextual references and are highly translated works. We provide a metric to characterize intertextuality at the corpus level and provide a quantitative analysis of the preservation of this rhetorical device across extant human translations and machine-generated counterparts. We go on to provide qualitative analysis of cases wherein human translations over- or underemphasize the intertextuality present in the text, whereas machine translations provide a neutral baseline. This provides support for established scholarship proposing that human translators have a propensity to amplify certain literary characteristics of the original manuscripts.

Large language models (LLMs) have exhibited impressive capabilities across a myriad of tasks, yet they occasionally yield undesirable outputs. We posit that these limitations are rooted in the foundational autoregressive architecture of LLMs, which inherently lacks mechanisms for differentiating between desirable and undesirable results. Drawing inspiration from the dual-process theory of human cognition, we introduce LLM2, a novel framework that combines an LLM (System 1) with a process-based verifier (System 2). Within LLM2, the LLM is responsible for generating plausible candidates, while the verifier provides timely process-based feedback to distinguish desirable and undesirable outputs. The verifier is trained with a pairwise comparison loss on synthetic process-supervision data generated through our token quality exploration strategy. Empirical results on mathematical reasoning benchmarks substantiate the efficacy of LLM2, exemplified by an accuracy enhancement from 50.3 to 57.8 (+7.5) for Llama3-1B on GSM8K. Furthermore, when combined with self-consistency, LLM2 achieves additional improvements, boosting major@20 accuracy from 56.2 to 70.2 (+14.0).

There has recently been considerable interest in incorporating information retrieval into large language models (LLMs). Retrieval from a dynamically expanding external corpus of text allows a model to incorporate current events and can be viewed as a form of episodic memory. Here we demonstrate that pre-processing the external corpus into semi-structured “atomic facts” makes retrieval more efficient. More specifically, we demonstrate that our particular form of atomic facts improves performance on various question answering tasks when the amount of retrieved text is limited. Limiting the amount of retrieval reduces the size of the context and improves inference efficiency.

Large Language Models (LLMs) have been shown to be biased in prior work, as they generate text that is in line with stereotypical views of the world or that is not representative of the viewpoints and values of historically marginalized demographic groups. In this work, we propose using data from parallel men’s and women’s events at the Olympic Games to investigate different forms of gender bias in language models. We define three metrics to measure bias, and find that models are consistently biased against women when the gender is ambiguous in the prompt. In this case, the model frequently retrieves only the results of the men’s event with or without acknowledging them as such, revealing pervasive gender bias in LLMs in the context of athletics.

Current machine translation (MT) systems for low-resource languages have a particular failure mode: When translating words in a given domain, they tend to confuse words within that domain. So, for example, “lion” might be translated as “alligator”, and “orange” might be rendered as “purple.” We propose a recall-based metric for measuring this problem and show that the problem exists in 122 low-resource languages. We then show that this problem can be mitigated by using a large language model (LLM) to post-edit the MT output, specifically by including the entire GATITOS lexicon for the relevant language as a very long context prompt. We show gains in average ChrF score over the set of 122 languages, and we show that the recall score for relevant lexical items also improves. Finally, we demonstrate that a small dedicated MT system with a general-purpose LLM as a post-editor is outperforms a lexicon-based RAG-LLM translator, suggesting a new paradigm for LLM use.

Large language models (LLMs) have become standard for natural language generation tasks, with instruction-tuning enhancing their capabilities. However, the lack of instruction-tuning datasets in languages other than English limits their application to diverse languages. To address this, researchers have adapted English-centric LLMs to other languages by appending English tuning data with its translated pair, from which we observe negative interference between the two. To resolve this, our contribution is identifying English as an internal pivot language, based on which we disentangle the roles of English and target language data in training. Specifically, we first design two roles as pivoted objectives, and also propose to regulate between the two, to better generalize for under-represented languages. Experiments across various languages demonstrate the effectiveness of our approach on multiple benchmarks. The code is publicly available for further exploration.

While Large Language Models (LLMs) have showcased remarkable proficiency in reasoning, there is still a concern about hallucinations and unreliable reasoning issues due to semantic associations and superficial logical chains. To evaluate the extent to which LLMs perform robust reasoning instead of relying on superficial logical chains, we propose a new evaluation dataset, the Concept-Reversed Winograd Schema Challenge (CR-WSC), based on the famous Winograd Schema Challenge (WSC) dataset. By simply reversing the concepts to those that are more associated with the wrong answer, we find that the performance of LLMs drops significantly despite the rationale of reasoning remaining the same. Furthermore, we propose Abstraction-of-Thought (AoT), a novel prompt method for recovering adversarial cases to normal cases using conceptual abstraction to improve LLMs’ robustness and consistency in reasoning, as demonstrated by experiments on CR-WSC.

Large Language Models (LLMs) have emerged as a dominant approach for a wide range of NLP tasks, with their access to external information further enhancing their capabilities. However, this introduces new vulnerabilities, known as prompt injection attacks, where external content embeds malicious instructions that manipulate the LLM’s output. Recently, the Base64 defense has been recognized as one of the most effective methods for reducing success rate of prompt injection attacks. Despite its efficacy, this method can degrade LLM performance on certain NLP tasks. To address this challenge, we propose a novel defense mechanism: mixture of encodings, which utilizes multiple character encodings, including Base64. Extensive experimental results show that our method achieves one of the lowest attack success rates under prompt injection attacks, while maintaining high performance across all NLP tasks, outperforming existing character encoding-based defense methods. This underscores the effectiveness of our mixture of encodings strategy for both safety and task performance metrics.

AI Safety Moderation (ASM) classifiers are designed to moderate content on social media platforms and to serve as guardrails that prevent Large Language Models (LLMs) from being fine-tuned on unsafe inputs. Owing to their potential for disparate impact, it is crucial to ensure that these classifiers: (1) do not unfairly classify content belonging to users from minority groups as unsafe compared to those from majority groups and (2) that their behavior remains robust and consistent across similar inputs. In this work, we thus examine the fairness and robustness of four widely-used, closed-source ASM classifiers: OpenAI Moderation API, Perspective API, Google Cloud Natural Language (GCNL) API, and Clarifai API. We assess fairness using metrics such as demographic parity and conditional statistical parity, comparing their performance against ASM models and a fair-only baseline. Additionally, we analyze robustness by testing the classifiers’ sensitivity to small and natural input perturbations. Our findings reveal potential fairness and robustness gaps, highlighting the need to mitigate these issues in future versions of these models.

Retrieval-Augmented Generation (RAG) models excel in knowledge-intensive tasks, especially under few-shot learning constraints. We introduce CoRAG, a framework extending RAG to collaborative settings, where clients jointly train a shared model using a collaborative passage store. To evaluate CoRAG, we introduce CRAB, a benchmark for collaborative homogeneous open-domain question answering. Our experiments demonstrate that CoRAG consistently outperforms both parametric collaborative learning methods and locally trained RAG models in low-resource scenarios. Further analysis reveals the critical importance of relevant passages within the shared store, the surprising benefits of incorporating irrelevant passages, and the potential for hard negatives to negatively impact performance. This introduces a novel consideration in collaborative RAG: the trade-off between leveraging a collectively enriched knowledge base and the potential risk of incorporating detrimental passages from other clients. Our findings underscore the viability of CoRAG, while also highlighting key design challenges and promising avenues for future research.

Endangered languages, such as Navajo—the most widely spoken Native American language—are significantly underrepresented in contemporary language technologies, exacerbating the challenges of their preservation and revitalization. This study evaluates Google’s Language Identification (LangID) tool, which does not currently support any Native American languages. To address this, we introduce a random forest classifier trained on Navajo and twenty erroneously suggested languages by LangID. Despite its simplicity, the classifier achieves near-perfect accuracy (97-100%). Additionally, the model demonstrates robustness across other Athabaskan languages—a family of Native American languages spoken primarily in Alaska, the Pacific Northwest, and parts of the Southwestern United States—suggesting its potential for broader application. Our findings underscore the pressing need for NLP systems that prioritize linguistic diversity and adaptability over centralized, one-size-fits-all solutions, especially in supporting underrepresented languages in a multicultural world. This work directly contributes to ongoing efforts to address cultural biases in language models and advocates for the development of culturally localized NLP tools that serve diverse linguistic communities.

The common practice of preprocessing text before feeding it into NLP models introduces many decision points which have unintended consequences on model performance. In this opinion piece, we focus on the handling of diacritics in texts originating in many languages and scripts. We demonstrate, through several case studies, the adverse effects of inconsistent encoding of diacritized characters and of removing diacritics altogether. We call on the community to adopt simple but necessary steps across all models and toolkits in order to improve handling of diacritized text and, by extension, increase equity in multilingual NLP.

Developing video captioning models is computationally expensive. The dynamic nature of video also complicates the design of multimodal models that can effectively caption these sequences. However, we find that by using minimal computational resources and without complex modifications to address video dynamics, an image-based model can be repurposed to outperform several specialised video captioning systems. Our adapted model demonstrates top-tier performance on major benchmarks, ranking 2nd on MSR-VTT and MSVD, and 3rd on VATEX. We transform it into a competitive video captioner by post-training a typical image captioning model BLIP-2 with only 6,000 video-text pairs and simply concatenating frames—significantly fewer data than other methods, which use 2.5 to 144 million pairs. From a resource optimization perspective, this video captioning study focuses on three fundamental factors: optimizing model scale, maximizing data efficiency, and incorporating reinforcement learning. This extensive study demonstrates that a lightweight, image-based adaptation strategy can rival state-of-the-art video captioning systems, offering a practical solution for low-resource scenarios.

Humans are accustomed to reading and writing in a forward manner, and this natural bias extends to text understanding in auto-regressive large language models (LLMs). This paper investigates whether LLMs, like humans, struggle with reverse modeling, specifically with reversed text inputs. We found that publicly available pre-trained LLMs cannot understand such inputs. However, LLMs trained from scratch with both forward and reverse texts can understand them equally well during inference across multiple languages.Our case study shows that different-content texts result in different losses if input (to LLMs) in different directions—some get lower losses for forward while some for reverse. This leads us to a simple and nice solution for data selection based on the loss differences between forward and reverse directions. Using our selected data in continued pretraining can boost LLMs’ performance by a large margin across different language understanding benchmarks.

A query encoder of a dual passage retrieval system can be tuned for specific types of queries or domains, while the precomputed and stored documents representations are kept intact. Switching from one query encoder to another when needed is easily feasible, unlike overhauling the embeddings of a whole knowledge base. In this work we raise a question: Can the generic, original qualities of the encoder be preserved or at least left not too degraded when it is tuned on a narrow domain? We conducted experiments on a high quality multilingual embedding model: Tuning it on a single English-only dataset, we observe that the tuning not only preserves the multilingual qualities, but even improves them. The embedding qualities on distinctly different data are also improved or at least preserved. Drawing on our observations, we suggest a more general hypothesis: Tuning with intentionally low learning rate can preserve or improve a system’s properties acquired in training, but not specifically targeted by tuning. We call this adiabatic tuning and provide tentative explanations.

A language can have different varieties. These varieties can affect the performance of natural language processing (NLP) models, including large language models (LLMs), which are often trained on data from widely spoken varieties. This paper introduces a novel and cost-effective approach to benchmark model performance across language varieties. We argue that international online review platforms,such as Booking.com, can serve as effective data sources for constructing datasets that capture comments in different language varieties from similar real-world scenarios, like reviews for the same hotel with the same rating using the same language (e.g., Mandarin Chinese) but different language varieties (e.g., Taiwan Mandarin, Mainland Mandarin). To prove this concept, we constructed a contextually aligned dataset comprising reviews in Taiwan Mandarin and Mainland Mandarin and tested six LLMs in a sentiment analysis task. Our results show that LLMs consistently underperform in Taiwan Mandarin.

Low-rank adaptation (LoRA) offers an efficient alternative to full-weight adaptation in federated fine-tuning of language models, significantly reducing computational costs. By adjusting ranks for each client, federated LoRA enables flexible resource allocation. However, we observe that heterogeneous ranks among clients lead to unstable performance. Our analysis attributes this instability to the conventional zero-padding aggregation strategy, which dilutes information from high-rank clients during model aggregation. To address this issue, we propose a replication-based padding strategy that better retains valuable information from clients with high-quality data. Empirically, this approach accelerates convergence and enhances the global model’s predictive performance.

Pre-training large language models (LLMs) faces significant memory challenges due to the large size of model weights. We introduce STaged parameter-Efficient Pre-training (STEP), which integrates parameter-efficient tuning techniques with model growth. We conduct experiments on pre-training LLMs of various sizes and demonstrate that STEP achieves up to a 53.9% reduction in maximum memory requirements compared to vanilla pre-training while maintaining equivalent performance. Furthermore, we show that the model by STEP performs comparably to vanilla pre-trained models on downstream tasks after instruction tuning.

Large language models (LLMs) frequently make errors when handling even simple numerical problems, such as comparing two small numbers. A natural hypothesis is that these errors stem from how LLMs represent numbers, and specifically, whether their representations of numbers capture their numeric values. We tackle this question from the observation that LLM errors on numerical tasks are often distributed across the digits of the answer rather than normally around its numeric value. Through a series of probing experiments and causal interventions, we show that LLMs internally represent numbers with individual circular representations per-digit in base 10.This digit-wise representation, as opposed to a value representation, sheds light on the error patterns of models on tasks involving numerical reasoning and could serve as a basis for future studies on analyzing numerical mechanisms in LLMs.

Recently, sharing key-value (KV) cache across layers has been found effective in efficient inference of large language models (LLMs). To systematically investigate different techniques of cross-layer KV sharing, we propose a unified framework that covers several recent methods and their novel variants. We conduct comprehensive experiments on all the configurations of the framework, evaluating their generation throughput and performance in language modeling and downstream tasks. We find that when reducing the size of the KV cache by 2×, most configurations can achieve higher throughput than standard transformers while maintaining competitive performance.When further reducing the size of the KV cache, however, pairing queries of all layers with KVs of upper layers performs better, at the expense of additional training cost and prefilling latency. We hope that this work will help users make more informed choices of cross-layer KV sharing approaches and facilitate future research on efficient LLM inference.

Spontaneous or conversational multilingual speech presents many challenges for state-of-the-art automatic speech recognition (ASR) systems. In this work, we present a new technique AMPS, that augments a multilingual multimodal ASR system with paraphrase-based supervision for improved conversational ASR in multiple languages, including Hindi, Marathi, Malayalam, Kannada, and Nyanja. We use paraphrases of the reference transcriptions as additional supervision while training the multimodal ASR model and selectively invoke this paraphrase objective for utterances with poor ASR performance. Using AMPS with a state-of-the-art multimodal model SeamlessM4T, we obtain significant relative reductions in word error rates (WERs) of up to 5%. We present detailed analyses of our system using both objective and human evaluation metrics.

While broad-coverage multilingual natural language processing tools have been developed, a significant portion of the world’s over 7000 languages are still neglected. One reason is the lack of evaluation datasets that cover a diverse range of languages, particularly those that are low-resource or endangered. To address this gap, we present a large-scale text classification dataset encompassing 1504 languages many of which have otherwise limited or no annotated data. This dataset is constructed using parallel translations of the Bible. We develop relevant topics, annotate the English data through crowdsourcing and project these annotations onto other languages via aligned verses. We benchmark a range of existing multilingual models on this dataset. We make our dataset and code available to the public.

The prevalence of toxic behavior in online gaming communities necessitates robust detection methods to ensure user safety. We introduce GameTox, a novel dataset comprising 53K game chat utterances annotated for toxicity detection through intent classification and slot filling. This dataset captures the complex relationship between user intent and specific linguistic features that contribute to toxic interactions. We extensively analyze the dataset to uncover key insights into the nature of toxic speech in gaming environments. Furthermore, we establish baseline performance metrics using state-of-the-art natural language processing and large language models, demonstrating the dataset’s contribution towards enhancing the detection of toxic behavior and revealing the limitations of contemporary models. Our results indicate that leveraging both intent detection and slot filling provides a significantly more granular and context-aware understanding of harmful messages. This dataset serves as a valuable resource to train advanced models that can effectively mitigate toxicity in online gaming and foster healthier digital spaces. Our dataset is publicly available at: https://github.com/shucoll/GameTox.

Summarization is one of the most common tasks performed by large language models (LLMs), especially in applications like Retrieval-Augmented Generation (RAG). However, existing evaluations of hallucinations in LLM-generated summaries, and evaluations of hallucination detection models both suffer from a lack of diversity and recency in the LLM and LLM families considered. This paper introduces FaithBench, a summarization hallucination benchmark comprising challenging hallucinations made by 10 modern LLMs from 8 different families, with ground truth annotations by human experts. “Challenging” here means summaries on which popular, state-of-the-art hallucination detection models, including GPT-4o-as-a-judge, disagreed on. Our results show GPT-4o and GPT-3.5-Turbo produce the least hallucinations. However, most state-of-the-art hallucination detection models have near 50% accuracies on FaithBench, indicating lots of room for future improvement.

The use of AI in legal analysis and prediction (LegalAI) has gained attention, with past research focusing on retrieval-based methods and fine-tuning large models. However, these approaches often require large datasets and underutilize the capabilities of modern large language models (LLMs). In this paper, inspired by the debate phase of real courtroom trials, we propose a novel legal judgment prediction model based on the Debate-Feedback architecture, which integrates LLM multi-agent debate and reliability evaluation models. Unlike traditional methods, our model achieves significant improvements in efficiency by minimizing the need for large historical datasets, thus offering a lightweight yet robust solution. Comparative experiments show that it outperforms several general-purpose and domain-specific legal models, offering a dynamic reasoning process and a promising direction for future LegalAI research.

K-nearest neighbor language models (kNN-LMs), which integrate retrieval with next-word prediction, have demonstrated strong performance in language modeling as well as some downstream NLP benchmarks. These results have led researchers to argue that models trained on poor quality or outdated data could perform well by employing a kNN extension that has access to a higher-quality datastore. In this work, we ask whether this improved ability to recall information really translates into downstream abilities. We extensively evaluate kNN-LMs on a diverse set of tasks, ranging from sentiment classification and commonsense reasoning to multi-hop reasoning. Results show that kNN-LMs excel at memory-intensive tasks, where utilizing the patterns in the input is sufficient for determining the output, but struggle with reasoning tasks that require integrating multiple pieces of information to derive new knowledge. We further demonstrate through oracle experiments and qualitative analysis that even with perfect retrieval, kNN-LMs still fail to determine the correct answers, placing an upper bound on their reasoning performance.

This paper introduces repetition neurons, which can be regarded as “skill neurons” responsible for the repetition problem in text generation tasks. These neurons are progressively activated more strongly as repetition continues, indicating that they perceive repetition as a task to copy the previous context repeatedly, similar to in-context learning. We identify these repetition neurons by comparing activation values before and after the onset of repetition in texts generated by recent pre-trained language models. We analyze the repetition neurons in three English and one Japanese pre-trained language models and observe similar patterns across them.

Model merging is an efficient way of obtaining a multi-task model from several pretrained models without further fine-tuning, and it has gained attention in various domains, including natural language processing (NLP). Despite the efficiency, a key challenge in model merging is the seemingly inevitable decrease in task performance as the number of models increases. In this paper, we propose **S**pectral **T**runcation **A**nd **R**escale (STAR) that aims at mitigating “merging conflicts” by truncating small components in the respective spectral spaces, which is followed by an automatic parameter rescaling scheme to retain the nuclear norm of the original matrix. STAR requires no additional inference on original training data and is robust to hyperparamater choice. We demonstrate the effectiveness of STAR through extensive model merging cases on diverse NLP tasks. Specifically, STAR works robustly across varying model sizes, and can outperform baselines by 4.2% when merging 12 models on Flan-T5. Our code is publicly available at https://github.com/IBM/STAR.

Modern language models (LMs) have significantly advanced generative modeling in natural language processing (NLP). Despite their success, LMs often struggle with adaptation to new contexts in real-time applications. A promising approach to task adaptation is activation intervention, which steers the LMs’ generation process by identifying and manipulating the activations. However, existing interventions rely heavily on heuristic rules or require many prompt inputs to determine effective interventions. In this paper, we propose a layer-wise additive activation intervention framework that optimizes the intervention process, thereby enhancing sample efficiency. We evaluate our framework on various datasets, demonstrating improvements in the accuracy of pretrained LMs and competing intervention baselines.

Automatically summarizing large text collections is a valuable tool for document research, with applications in journalism, academic research, legal work, and many other fields. In this work, we contrast two classes of systems for large-scale multi-document summarization (MDS): compression and full-text. Compression-based methods use a multi-stage pipeline and often lead to lossy summaries. Full-text methods promise a lossless summary by relying on recent advances in long-context reasoning. To understand their utility on large-scale MDS, we evaluated them on three datasets, each containing approximately one hundred documents per summary. Our experiments cover a diverse set of long-context transformers (Llama-3.1, Command-R, Jamba-1.5-Mini) and compression methods (retrieval-augmented, hierarchical, incremental). Overall, we find that full-text and retrieval methods perform the best in most settings. With further analysis into the salient information retention patterns, we show that compression-based methods show strong promise at intermediate stages, even outperforming full-context. However, they suffer information loss due to their multi-stage pipeline and lack of global context. Our results highlight the need to develop hybrid approaches that combine compression and full-text approaches for optimal performance on large-scale multi-document summarization.

Large Vision Language Models (LVLMs) often suffer from object hallucination, which undermines their reliability. Surprisingly, we find that simple object-based visual prompting—overlaying visual cues (e.g., bounding box, circle) on images—can significantly mitigate such hallucination; however, different visual prompts (VPs) vary in effectiveness. To address this, we propose Black-Box Visual Prompt Engineering (BBVPE), a framework to identify optimal VPs that enhance LVLM responses without needing access to model internals. Our approach employs a pool of candidate VPs and trains a router model to dynamically select the most effective VP for a given input image. This black-box approach is model-agnostic, making it applicable to both open-source and proprietary LVLMs. Evaluations on benchmarks such as POPE and CHAIR demonstrate that BBVPE effectively reduces object hallucination.

Distinguishing between extremely similar diseases is a critical and challenging aspect of clinical decision-making. Traditional classification, contrastive learning, and Large Language Models (LLMs) based methods fail to detect the subtle clues necessary for differentiation. This task demands complex reasoning and a variety of tools to identify minor differences and make informed decisions. This paper probes a novel framework that leverages LLMs and a multi-agent system to achieve accurate disease diagnosis through a process of repeated debate and reassessment. The approach aims to identify subtle differences between similar disease candidates. We structure patient information and integrate extensive medical knowledge to guide the analysis towards discerning these differences for precise diagnosis. Comprehensive experiments were conducted on two public datasets and two newly introduced datasets, JarvisD2-Chinese and JarvisD2-English, to validate the effectiveness of our method. The results confirm the efficacy of our approach, demonstrating its potential to enhance diagnostic precision in healthcare.

This paper investigates whether large language models (LLMs) utilize numerical attributes encoded in a low-dimensional subspace of theembedding space when answering questions involving numeric comparisons, e.g., Was Cristiano born before Messi? We first identified,using partial least squares regression, these subspaces, which effectively encode the numerical attributes associated with the entities in comparison prompts. Further, we demonstrate causality, by intervening in these subspaces to manipulate hidden states, thereby altering the LLM’s comparison outcomes. Experiments conducted on three different LLMs showed that our results hold across different numerical attributes, indicating that LLMs utilize the linearly encoded information for numerical reasoning.

Realignment techniques are often employed to enhance cross-lingual transfer in multilingual language models, still, they can sometimes degrade performance in languages that differ significantly from the fine-tuned source language. This paper introduces AlignFreeze, a method that freezes either the layers’ lower half or upper half during realignment. Through controlled experiments on 4 tasks, 3 models, and in 35 languages, we find that realignment affects all the layers but can be the most detrimental to the lower ones. Freezing the lower layers can prevent performance degradation. Particularly, AlignFreeze improves Part-of-Speech (PoS) tagging performances in languages where full realignment fails: with XLM-R, it provides improvements of more than one standard deviation in accuracy in seven more languages than full realignment.

Tracking a patient’s cognitive status early in the onset of the disease provides an opportunity to diagnose and intervene in Alzheimer’s disease (AD). However, relying solely on magnetic resonance imaging (MRI) images with traditional classification and regression models may not fully extract finer-grained information. This study proposes a multi-task Fusion Language Image Question Answering model (FLIQA-AD) to perform AD identification and Mini Mental State Examination (MMSE) prediction. Specifically, a 3D Adapter is introduced in Vision Transformer (ViT) model for image feature extraction. The patient electronic health records (EHR) information and questions related to the disease work as text prompts to be encoded. Then, an ADFormer model, which combines self-attention and cross-attention mechanisms, is used to capture the correlation between EHR information and structure features. After that, the extracted brain structural information and textual content are combined as input sequences for the large language model (LLM) to identify AD and predict the corresponding MMSE score. Experimental results demonstrate the strong discrimination and MMSE prediction performance of the model, as well as question-answer capabilities.

In this paper, we introduce Transform Retrieval, a novel approach aimed at improving Textual Entailment Retrieval within the framework of Retrieval-Augmented Generation (RAG). While RAG has shown promise in enhancing Large Language Models by retrieving relevant documents to extract specific knowledge or mitigate hallucination, current retrieval methods often prioritize relevance without ensuring the retrieved documents semantically support answering the queries. Transform Retrieval addresses this gap by transforming query embeddings to better align with semantic entailment without re-encoding the document corpus. We achieve this by using a transform model and employing a contrastive learning strategy to optimize the alignment between transformed query embeddings and document embeddings for better entailment.We evaluated the framework using BERT as frozen pre-trained encoder and compared it with a fully fine-tuned skyline model. Experimental results show that Transform Retrieval with simple MLP consistently approaches the skyline across multiple datasets, demonstrating the method’s effectiveness. The high performance on HotpotQA highlights its strength in many-to-many retrieval scenarios.

Multimodal foundation models aim to create a unified representation space that abstracts away from surface features like language syntax or modality differences. To investigate this, we study the internal representations of three recent models, analyzing the model activations from semantically equivalent sentences across languages in the text and speech modalities. Our findings reveal that: 1) Cross-modal representations converge over model layers, except in the initial layers specialized at text and speech processing. 2) Length adaptation is crucial for reducing the cross-modal gap between text and speech, although current approaches’ effectiveness is primarily limited to high-resource languages. 3) Speech exhibits larger cross-lingual differences than text. 4) For models not explicitly trained for modality-agnostic representations, the modality gap is more prominent than the language gap.

Reasoning is a central capability of human intelligence. In recent years, with the advent of large-scale datasets, pretrained large language models have emerged with new capabilities, including reasoning. However, these models still struggle with long-term, complex reasoning tasks, such as playing chess. Based on the observation that expert chess players employ a dual approach combining long-term strategic play with short-term tactical play along with language explanation, we propose improving the reasoning capability of large language models in chess by integrating annotated strategy and tactic. Specifically, we collect a dataset named MATE, which consists of 1 million chess positions with candidate moves annotated for strategy and tactics. We finetune the LLaMA-3-8B model and compare it against state-of-the-art commercial language models in the task of selecting better chess moves. Our experiments show that our models perform better than GPT, Claude, and Gemini models. We find that language explanations can enhance the reasoning capability of large language models.

Graph databases like Neo4j are gaining popularity for handling complex, interconnected data, over traditional relational databases in modeling and querying relationships. While translating natural language into SQL queries is well-researched, generating Cypher queries for Neo4j remains relatively underexplored. In this work, we present an automated, LLM Supervised, pipeline to generate high quality synthetic data for Text2Cypher. Our Cypher data generation pipeline introduces LLM-As-Database-Filler, a novel strategy for ensuring Cypher query correctness, thus resulting in high quality generations. Using our pipeline, we generate high quality Text2Cypher data - SynthCypher containing 29.8k instances across various domains and queries with varying complexities. Training open-source LLMs like LLaMa-3.1-8B, Mistral-7B, and QWEN7B on SynthCypher results in performance gains of up to 40% on the Text2Cypher test split and 30% on the SPIDER benchmark, adapted for graph databases.

Natural Language Inference (NLI) is crucial for evaluating models’ Natural Language Understanding (NLU) and reasoning abilities. The development of NLI, in part, has been driven by the creation of large datasets, which require significant human effort. This has spurred interest in semi-supervised learning (SSL) that leverages both labeled and unlabeled data. However, the absence of hypotheses and class labels in NLI tasks complicates SSL. Prior work has used class-specific fine-tuned large language models (LLMs) to generate hypotheses and assign pseudo-labels but discarded many LLM-constructed samples during training to ensure the quality. In contrast, we propose to leverage all LLM-constructed samples by handling potentially noisy samples by injecting the moments of labeled samples during training to properly adjust the level of noise. Our method outperforms strong baselines on multiple NLI datasets in low-resource settings.

Most large language models are multilingual instruction executors. Prior studies suggested that English instructions are more effective than target-language instructions even for non-English tasks; however, these studies often use datasets and instructions translated from English, which introduce biases known as translationese, hindering an unbiased comparison. To address this issue, we conduct a fair comparison between English and target-language instructions by eliminating translationese effects. Contrary to previous studies, our experiments across several tasks reveal that the advantage of adopting English instructions is not overwhelming. Additionally, we report on the features of generated texts and the instruction-following abilities when using respective instructions.

This paper presents the Korean National Educational Test Benchmark (KoNET), a new benchmark designed to evaluate Multimodal Generative AI Systems using Korean national educational tests. KoNET comprises four exams: the Korean Elementary General Educational Development Test (KoEGED), Middle (KoMGED), High (KoHGED), and College Scholastic Ability Test (KoCSAT). These exams are renowned for their rigorous standards and diverse questions, facilitating a comprehensive analysis of AI performance across different educational levels. By focusing on Korean, KoNET provides insights into model performance in less-explored languages. We assess a range of models—open-source, open-access, and closed APIs—by examining difficulties, subject diversity, and human error rates. The code and dataset builder will be made fully open-source.

Recent advancements in large multimodal models (LMMs) have showcased impressive code generation capabilities, primarily evaluated through image-to-code benchmarks. However, these benchmarks are limited to specific visual programming scenarios where the logic reasoning and the multimodal understanding capacities are split apart. To fill this gap, we propose ScratchEval, a novel benchmark designed to evaluate the visual programming reasoning ability of LMMs. ScratchEval is based on Scratch, a block-based visual programming language widely used in children’s programming education. By integrating visual elements and embedded programming logic, ScratchEval requires the model to process both visual information and code structure, thereby comprehensively evaluating its programming intent understanding ability. Our evaluation approach goes beyond the traditional image-to-code mapping and focuses on unified logical thinking and problem-solving abilities, providing a more comprehensive and challenging framework for evaluating the visual programming ability of LMMs. ScratchEval not only fills the gap in existing evaluation methods, but also provides new insights for the future development of LMMs in the field of visual programming.

Dense embeddings deliver strong retrieval performance but often lack interpretability and controllability. This paper introduces a novel approach using sparse autoencoders (SAE) to interpret and control dense embeddings via the learned latent sparse features. Our key contribution is the development of a retrieval-oriented contrastive loss, which ensures the sparse latent features remain effective for retrieval tasks and thus meaningful to interpret. Experimental results demonstrate that both the learned latent sparse features and their reconstructed embeddings retain nearly the same retrieval accuracy as the original dense vectors, affirming their faithfulness. Our further examination of the sparse latent space reveals interesting features underlying the dense embeddings and we can control the retrieval behaviors via manipulating the latent sparse features, for example, prioritizing documents from specific perspectives in the retrieval results.

As large language models (LLMs) generate more human-like texts, concerns about the side effects of AI-generated texts (AIGT) have grown. So, researchers have developed methods for detecting AIGT. However, two challenges remain. First, the performance of detecting black-box LLMs is low because existing models focus on probabilistic features. Second, most AIGT detectors have been tested on a single-candidate setting, which assumes that we know the origin of an AIGT and which may deviate from the real-world scenario. To resolve these challenges, we propose DART, which consists of four steps: rephrasing, semantic parsing, scoring, and multiclass classification. We conducted three experiments to test the performance of DART. The experimental result shows that DART can discriminate multiple black-box LLMs without probabilistic features and the origin of AIGT.

We propose a novel architecture for graph-based dependency parsing that explicitly constructs vectors, from which both arcs and labels are scored. Our method addresses key limitations of the standard two-pipeline approach by unifying arc scoring and labeling into a single network, reducing scalability issues caused by the information bottleneck and lack of parameter sharing. Additionally, our architecture overcomes limited arc interactions with transformer layers to efficiently simulate higher-order dependencies. Experiments on PTB and UD show that our model outperforms state-of-the-art parsers in both accuracy and efficiency.

We examine the pre-training dynamics of language models, focusing on their ability to copy text from preceding context—a fundamental skill for various LLM applications, including in-context learning (ICL) and retrieval-augmented generation (RAG). We propose a novel perspective that Transformer-based language models develop copying abilities similarly to grokking, which refers to sudden generalization on test set long after the model fit to the training set. Our experiments yield three arguments: (1) The pre-training loss decreases rapidly, while the context copying ability of models initially lags and then abruptly saturates. (2) The speed of developing copying ability is independent of the number of tokens trained, similarly to how grokking speed is unaffected by dataset size as long as the data distribution is preserved. (3) Induction heads, the attention heads responsible for copying, form from shallow to deep layers during training, mirroring the development of circuits in deeper layers during grokking. We contend that the connection between grokking and context copying can provide valuable insights for more effective language model training, ultimately improving in-context performance. For example, we demonstrated that techniques that enhance grokking, such as regularization, either accelerate or enhance the development of context copying.

Large Language Models (LLMs) have shown promising results in a variety of literary tasks, often using complex memorized details of narration and fictional characters. In this work, we evaluate the ability of Llama-3 at attributing utterances of direct-speech to their speaker in novels. The LLM shows impressive results on a corpus of 28 novels, surpassing published results with ChatGPT and encoder-based baselines by a large margin. We then validate these results by assessing the impact of book memorization and annotation contamination.We found that these types of memorization do not explain the large performance gain, making Llama-3 the new state-of-the-art for quotation attribution in English literature. We release publicly our code and data.

Previous work has considered token overlap, or even similarity of token distributions, as predictors for multilinguality and cross-lingual knowledge transfer in language models. However, these very literal metrics assign large distances to language pairs with different scripts, which can nevertheless show good cross-linguality. This limits the explanatory strength of token overlap for knowledge transfer between language pairs that use distinct scripts or follow different orthographic conventions. In this paper, we propose subword token alignability as a new way to understand the impact and quality of multilingual tokenisation. In particular, this metric predicts multilinguality much better when scripts are disparate and the overlap of literal tokens is low. We analyse this metric in the context of both encoder and decoder models, look at data size as a potential distractor, and discuss how this insight may be applied to multilingual tokenisation in future work. We recommend our subword token alignability metric for identifying optimal language pairs for cross-lingual transfer, as well as to guide the construction of better multilingual tokenisers in the future. We publish our code and reproducibility details.

Recent evaluations of LLMs on coreference resolution have revealed that traditional output formats and evaluation metrics do not fully capture the models’ referential understanding. To address this, we introduce IdentifyMe, a new benchmark for mention resolution presented in a multiple-choice question (MCQ) format, commonly used for evaluating LLMs. IdentifyMe features long narratives and employs heuristics to exclude easily identifiable mentions, creating a more challenging task. The benchmark also consists of a curated mixture of different mention types and corresponding entities, allowing for a fine-grained model performance analysis. We evaluate both closed- and open-source LLMs on IdentifyMe and observe a significant performance gap (20-30%) between the state-of-the-art sub-10B open models vs. closed ones. We observe that pronominal mentions, which have limited surface information, are typically harder for models to resolve than nominal mentions. Additionally, we find that LLMs often confuse entities when their mentions overlap in nested structures. The highest scoring model, GPT-4o, achieves 81.9% accuracy, highlighting the strong referential capabilities of state-of-the-art LLMs while also indicating room for further improvement.

While recent zero-shot multi-speaker text-to-speech (TTS) models achieve impressive results, they typically rely on extensive transcribed speech datasets from numerous speakers and intricate training pipelines. Meanwhile, self-supervised learning (SSL) speech features have emerged as effective intermediate representations for TTS. Further, SSL features from different speakers that are linearly close share phonetic information while maintaining individual speaker identity. In this study, we introduce kNN-TTS, a simple and effective framework for zero-shot multi-speaker TTS using retrieval methods which leverage the linear relationships between SSL features. Objective and subjective evaluations show that our models, trained on transcribed speech from a single speaker only, achieve performance comparable to state-of-the-art models that are trained on significantly larger training datasets. The low training data requirements mean that kNN-TTS is well suited for the development of multi-speaker TTS systems for low-resource domains and languages. We also introduce an interpolation parameter which enables fine-grained voice morphing. Demo samples are available at https://idiap.github.io/knn-tts .

With the adoption of retrieval-augmented generation (RAG), large language models (LLMs) are expected to ground their generation to the retrieved contexts. Yet, this is hindered by position bias of LLMs, failing to evenly attend to all contexts. Previous work has addressed this by synthesizing contexts with perturbed positions of gold segment, creating a position-diversified train set. We extend this intuition to propose consistency regularization with augmentation and distillation. First, we augment each training instance with its position perturbation to encourage consistent predictions, regardless of ordering. We also distill behaviors of this pair, although it can be counterproductive in certain RAG scenarios where the given order from the retriever is crucial for generation quality. We thus propose CORD, balancing COnsistency and Rank Distillation: CORD adaptively samples noise-controlled perturbations from an interpolation space, ensuring both consistency and respect for the rank prior. Empirical results show this balance enables CORD to outperform consistently in diverse RAG benchmarks.

Heterogeneous graph neural networks have recently gained attention for long document summarization, modeling the extraction as a node classification task. Although effective, these models often require external tools or additional machine learning models to define graph components, producing highly complex and less intuitive structures. We present GraphLSS, a heterogeneous graph construction for long document extractive summarization, incorporating Lexical, Structural, and Semantic features. It defines two levels of information (words and sentences) and four types of edges (sentence semantic similarity, sentence occurrence order, word in sentence, and word semantic similarity) without any need for auxiliary learning models. Experiments on two benchmark datasets show that GraphLSS is competitive with top-performing graph-based methods, outperforming recent non-graph models. We release our code on GitHub.

Fact verification (FV) aims to assess the veracity of a claim based on relevant evidence. The traditional approach for automated FV includes a three-part pipeline relying on short evidence snippets and encoder-only inference models. More recent approaches leverage the multi-turn nature of LLMs to address FV as a step-by-step problem where questions inquiring additional context are generated and answered until there is enough information to make a decision. This iterative method makes the verification process rational and explainable. While these methods have been tested for encyclopedic claims, exploration on domain-specific and realistic claims is missing. In this work, we apply an iterative FV system on three medical fact-checking datasets and evaluate it with multiple settings, including different LLMs, external web search, and structured reasoning using logic predicates. We demonstrate improvements in the final performance over traditional approaches and the high potential of step-by-step FV systems for domain-specific claims.

We present lightweight flow matching multilingual text-to-speech (TTS) systems for Ojibwe, Mi’kmaq, and Maliseet, three Indigenous languages in North America. Our results show that training a multilingual TTS model on three typologically similar languages can improve the performance over monolingual models, especially when data are scarce. Attention-free architectures are highly competitive with self-attention architecture with higher memory efficiency. Our research provides technical development to language revitalization for low-resource languages but also highlights the cultural gap in human evaluation protocols, calling for a more community-centered approach to human evaluation.

Training conversational question-answering (QA) systems demands a substantial amount of in-domain data, which is often scarce in practice. A common solution to this challenge is to generate synthetic data. Traditional methods typically follow a top-down approach, where a large language model (LLM) generates multi-turn dialogues from a broad prompt. While this method produces coherent conversations, it offers limited fine-grained control over the content and is susceptible to hallucinations. We introduce a bottom-up conversation synthesis approach, where QA pairs are generated first and then combined into a coherent dialogue. This method offers greater control and precision by dividing the process into two distinct steps, enabling refined instructions and validations to be handled separately. Additionally, this structure allows the use of non-local models in stages that do not involve proprietary knowledge, enhancing the overall quality of the generated data. Both human and automated evaluations demonstrate that our approach produces more realistic and higher-quality dialogues compared to top-down methods.

Human judgments are inherently subjective and are actively affected by personal traits such as gender and ethnicity. While Large LanguageModels (LLMs) are widely used to simulate human responses across diverse contexts, their ability to account for demographic differencesin subjective tasks remains uncertain. In this study, leveraging the POPQUORN dataset, we evaluate nine popular LLMs on their abilityto understand demographic differences in two subjective judgment tasks: politeness and offensiveness. We find that in zero-shot settings, most models’ predictions for both tasks align more closely with labels from White participants than those from Asian or Black participants, while only a minor gender bias favoring women appears in the politeness task. Furthermore, sociodemographic prompting does not consistently improve and, in some cases, worsens LLMs’ ability to perceive language from specific sub-populations. These findings highlight potential demographic biases in LLMs when performing subjective judgment tasks and underscore the limitations of sociodemographic prompting as a strategy to achieve pluralistic alignment. Code and data are available at: https://github.com/Jiaxin-Pei/LLM-as-Subjective-Judge.

Large language models (LLMs) struggle in social science domains, where critical thinking and human-level inference are crucial. In this work, we propose a multi-agent social reasoning framework that leverages the generative and reasoning capabilities of LLMs to generate and evaluate reasons from multiple perspectives grounded in social science theories, and construct a factor graph for inference. Experimental results on understanding power dynamics in conversations show that our method outperforms standard prompting baselines, demonstrating its potential for tackling hard Computational Social Science (CSS) tasks.

Cultural and language factors significantly influence counseling, but Natural Language Processing research has not yet examined whether the findings of conversational analysis for counseling conducted in English apply to other languages. This paper presents a first step towards this direction. We introduce MIDAS (Motivational Interviewing Dataset in Spanish), a counseling dataset created from public video sources that contains expert annotations for counseling reflections and questions. Using this dataset, we explore language-based differences in counselor behavior in English and Spanish and develop classifiers in monolingual and multilingual settings, demonstrating its applications in counselor behavioral coding tasks.

Large language models (LLMs) have shown remarkable advances in language generation and understanding but are also prone to exhibiting harmful social biases. While recognition of these behaviors has generated an abundance of bias mitigation techniques, most require modifications to the training data, model parameters, or decoding strategy, which may be infeasible without access to a trainable model. In this work, we leverage the zero-shot capabilities of LLMs to reduce stereotyping in a technique we introduce as zero-shot self-debiasing. With two approaches, self-debiasing via explanation and self-debiasing via reprompting, we show that self-debiasing can significantly reduce the degree of stereotyping across nine different social groups while relying only on the LLM itself and a simple prompt, with explanations correctly identifying invalid assumptions and reprompting delivering the greatest reductions in bias. We hope this work opens inquiry into other zero-shot techniques for bias mitigation.

This study finds that existing information retrieval (IR) models show significant biases based on the linguistic complexity of input queries, performing well on linguistically simpler (or more complex) queries while underperforming on linguistically more complex (or simpler) queries.To address this issue, we propose EqualizeIR, a framework to mitigate linguistic biases in IR models. EqualizeIR uses a linguistically biased weak learner to capture linguistic biases in IR datasets and then trains a robust model by regularizing and refining its predictions using the biased weak learner. This approach effectively prevents the robust model from overfitting to specific linguistic patterns in data. We propose four approaches for developing linguistically-biased models. Extensive experiments on several datasets show that our method reduces performance disparities across linguistically simple and complex queries, while improving overall retrieval performance.

Open-vocabulary audio language models (ALMs), like Contrastive Language Audio Pretraining (CLAP), represent a promising new paradigm for audio-text retrieval using natural language queries. In this paper, for the first time, we perform controlled experiments on various benchmarks to show that existing ALMs struggle to generalize to linguistic variations in textual queries. To address this issue, we propose RobustCLAP, a novel and compute-efficient technique to learn audio-language representations agnostic to linguistic variations. Specifically, we reformulate the contrastive loss used in CLAP architectures by introducing a multi-view contrastive learning objective, where paraphrases are treated as different views of the same audio scene and use this for training. Our proposed approach improves the text-to-audio retrieval performance of CLAP by 0.8%-13% across benchmarks and enhances robustness to linguistic variation. We make our code publicly available

Automatic Post-Editing (APE) systems often struggle with over-correction, where unnecessary modifications are made to a translation, diverging from the principle of minimal editing. In this paper, we propose a novel technique to mitigate over-correction by incorporating word-level Quality Estimation (QE) information during the decoding process. This method is architecture-agnostic, making it adaptable to any APE system, regardless of the underlying model or training approach. Our experiments on English-German, English-Hindi, and English-Marathi language pairs show the proposed approach yields significant improvements over their corresponding baseline APE systems, with TER gains of 0.65, 1.86, and 1.44 points, respectively. These results underscore the complementary relationship between QE and APE tasks and highlight the effectiveness of integrating QE information to reduce over-correction in APE systems.

Large language models (LLMs) still lack delicate controllability over their responses, which is critical to enhancing their performance and the user experience. However, curating supervised fine-tuning (SFT) datasets to improve LLM controllability usually relies on human experts or proprietary LLMs, which requires additional costs. To bridge this gap, we propose Rule-based Data Recycling (RuleR), a data augmentation method incorporating multiple constraints into the original data samples according to predefined rules, which creates new training tasks to consolidate the controllability of LLMs. Instead of creating new data from scratch, RuleR “recycles” existing data by simply applying rule-based edits to their responses and appending the rule-instructions in their original instructions. Experimental results demonstrate RuleR’s effectiveness in improving LLM controllability while maintaining general instruction-following capabilities.

The growing use of large language models (LLMs) in academic peer review poses significant challenges, particularly in distinguishing AI-generated content from human-written feedback. This research addresses the problem of identifying AI-generated peer review comments, which are crucial to maintaining the integrity of scholarly evaluation. Prior research has primarily focused on generic AI-generated text detection or on estimating the fraction of peer reviews that may be AI-generated, often treating reviews as monolithic units. However, these methods fail to detect finer-grained AI-generated points within mixed-authorship reviews. To address this gap, we propose MixRevDetect, a novel method to identify AI-generated points in peer reviews. Our approach achieved an F1 score of 88.86%, significantly outperforming existing AI text detection methods.

Summarizing movie screenplays presents a unique set of challenges compared to standard document summarization. Screenplays are not only lengthy, but also feature a complex interplay of characters, dialogues, and scenes, with numerous direct and subtle relationships and contextual nuances that are difficult for machine learning models to accurately capture and comprehend. Recent attempts at screenplay summarization focus on fine-tuning transformer-based pre-trained models, but these models often fall short in capturing long-term dependencies and latent relationships, and frequently encounter the “lost in the middle” issue. To address these challenges, we introduce DiscoGraMS, a novel resource that represents movie scripts as a movie character-aware discourse graph (CaD Graph). This approach is well-suited for various downstream tasks, such as summarization, question-answering, and salience detection. The model aims to preserve all salient information, offering a more comprehensive and faithful representation of the screenplay’s content. We further explore a baseline method that combines the CaD Graph with the corresponding movie script through a late fusion of graph and text modalities, and we present very initial promising results. We have made our code and dataset publicly available.

While NLP models often seek to capture cognitive states via language, the validity of predicted states is determined by comparing them to annotations created without access the cognitive states of the authors. In behavioral sciences, cognitive states are instead measured via experiments. Here, we introduce an experiment-based framework for evaluating language-based cognitive style models against human behavior. We explore the phenomenon of decision making, and its relationship to the linguistic style of an individual talking about a recent decision they made. The participants then follow a classical decision-making experiment that captures their cognitive style, determined by how preferences change during a decision exercise. We find that language features, intended to capture cognitive style, can predict participants’ decision style with moderate-to-high accuracy (AUC 0.8), demonstrating that cognitive style can be partly captured and revealed by discourse patterns.