Retrieval-Augmented Generation (RAG) introduces a specific failure mode in discrete diffusion language models: when retrieved context contradicts parametric knowledge, the iterative denoising process becomes a visible battleground between competing knowledge sources. We identify temporal semantic divergence as an observable for detecting these conflicts and introduce the Trajectory Variance Score (TVS), a simple and interpretable measure of this divergence. TVS computes the mean pairwise cosine distance of answer embeddings across independent stochastic denoising trajectories, capturing the temporal tug of war between parametric and contextual attractors. Requiring as few as two parallel inference runs, TVS is computationally lightweight. Across four diverse datasets (Synthetic, SciQ, PopQA, and CounterFact), a simple Logistic Regression classifier using TVS achieves 70.10% accuracy and 0.7647 AUROC on LLaDA. On Dream 7B, increasing the number of trajectories from two to five improves accuracy from 63.91% to 69.62%. More complex sequential models provide only marginal improvements over the linear classifier. Evaluation across LLaDA and Dream 7B demonstrates that conflict-induced trajectory dynamics and their key properties transfer across distinct diffusion architectures.

Large language models (LLMs) are increasingly deployed for translation tasks, yet their implicit political positioning in such contexts remains understudied. We ask whether a single politically charged framing term, such as aggressor, enemy, neighbour, or coloniser is sufficient to trigger implicit political alignment in an otherwise apolitical task. We present a fully crossed factorial study in which eight models spanning Western, Chinese, and European origins are prompted to translate culturally attributed recipes into a target language left deliberately unspecified. Across 17 languages, four framing conditions, eight models, and 15,680 responses, we find that models do not simply decline or ask for clarification but resolve the ambiguity. Language resolution and reasoning behavior cluster meaningfully along model families: Western models hedge and deflect with vague justifications, Chinese models resolve conflicts silently, and Mistral Large emerges as a distinct profile combining high compliance with conflict-grounded reasoning. Sensitivity to framing terms is consistent across models: even subtle framing variation is sufficient to modulate behavior. Our findings urge caution when deploying LLMs for translation in conflict-adjacent contexts, where implicit political judgments may be made without any signal to the user.

Commonsense reasoning is a key language model capability, as it is purportedly a prerequisite for many basic tasks, unlike specific factual knowledge. It is often measured with multiple-choice questions (MCQ) benchmarks, e.g. HellaSwag and PIQA. Some of these benchmarks, however, are outdated and contain numerous validity issues. We illustrate some typical validity issues with a case study on HellaSwag, one of the most popular and problematic benchmarks for commonsense reasoning. The issues we find range from basic ungrammaticality and numerous typos to misleading prompts or equally correct options. We show that if we remove question prompts or replace them with "Lorem ipsum dolor...", about 68% of model predictions remain the same. We argue that this occurs due to inner flaws in the benchmark, not mere contamination that might be present in some models. Since benchmark scores are an essential part of model selection in both research and commercial applications, these validity issues can have severe consequences. Based on our findings, we propose BenCheck, a package for benchmark validity analysis that encapsulates the main checks performed in our case study and can be used to audit MCQ benchmarks. We apply these checks to other commonsense reasoning benchmarks.

Real-world personally identifiable information (PII) redaction often operates on document images—scans, screenshots, and PDF renderings—where OCR errors, layout structure, and visual noise determine whether sensitive information is actually removed. Existing PII benchmarks are mostly text-centric and do not measure document-level redaction risk: a page remains unsafe if even one identifier is missed. We introduce LeakageBench, a challenge set of 500 document images with 11,954 GDPR-aligned PII annotations spanning direct identifiers, linkage keys, and contextual re-identification surfaces. We evaluate generic OCR pipelines, commercial and task-adapted OCR-dependent detectors, and OCR-free vision-language models using entity-level F1, group-wise leakage, and document-level leakage metrics. Code Interpreter raises GPT-5.5 localization F1 from 0.090 to 0.249, but critical page-level leakage remains 0.968. These results show that stronger detection and tool assistance improve localization without making most pages safe for release. LeakageBench provides a diagnostic benchmark for high-recall, spatially grounded PII redaction in document images.

Tokenizer adaptation plays an important role in adapting pre-trained language models to new domains or languages. In this work, we address two complementary aspects of this process: vocabulary extension and pruning. The common approach to extension trains a new tokenizer on domain-specific text and appends the tokens that do not overlap with the existing vocabulary, which often results in many tokens that are unreachable or never used. We propose continued BPE training that extends a pre-trained tokenizer by continuing the BPE merge learning process on new data. Experiments across multiple languages and model families show that this approach improves tokenization efficiency and leads to better utilization of added vocabulary. We also introduce leaf-based vocabulary pruning, which removes redundant tokens while preserving model quality. Together, these methods provide practical tools for controlled vocabulary modification, which we release as an open-source toolkit.

Automated analysis of customer feedback on social media is hindered by three challenges: the high cost of annotated training data, the scarcity of evaluation sets, especially in multilingual settings, and privacy concerns that prevent data sharing and reproducibility. We address these issues by developing a generalizable synthetic data generation pipeline applied to a case study on customer distress detection in French public transportation. Our approach utilizes backtranslation with fine-tuned models to generate 1.7 million synthetic tweets from a small seed corpus, complemented by synthetic reasoning traces. We train 600M-parameter reasoners with English and French reasoning that achieve 77-79% accuracy on human-annotated evaluation data, matching or exceeding SOTA proprietary LLMs and specialized encoders. Beyond reducing annotation costs, our pipeline preserves privacy by eliminating the exposure of sensitive user data. Our methodology can be adopted for other use cases and languages.

LLMs often exhibit highly agreeable conversational styles, also known as AI sycophancy. This pattern may become problematic when interacting with user prompts that reflect negative social tendencies, risking the amplification of harmful behavior. We examine how LLMs respond to user prompts expressing varying degrees of Dark Triad traits (Machiavellianism, Narcissism, and Psychopathy) using a curated dataset. Our analysis reveals systematic differences across models: while all models predominantly exhibit corrective behavior, some generate reinforcing or ambivalent output. Model behavior further varies with severity level and response sentiment. These findings highlight the need for safer conversational systems that can reliably detect and respond to users escalating from benign to harmful requests.

RAG systems are usually compared as complete pipelines, making it difficult to attribute improvements to individual components. We evaluate chunking, retrieval, fusion, and reranking through a full factorial design with paired component substitutions and equal context budgets. Across five benchmarks, chunking interacts with retrieval architecture, BM25 fusion changes direction by domain, and reranking is the only intervention we test that remains beneficial across nearly all upstream configurations. Claim-level transitions further show that similar net recall can conceal very different changes in visible evidence. These findings demonstrate that RAG components cannot be ranked independently and motivate interaction-aware evaluation instead of winner-take-all pipeline comparisons.

Efficiency and safety of Large Language Models (LLMs), among other factors, rely on the quality of tokenization. A good tokenizer not only improves inference speed and language understanding but also provides extra defense against jailbreak attacks and lowers the risk of hallucinations. In this work, we investigate the efficiency of code tokenization, in particular from the perspective of data source diversity. We demonstrate that code tokenizers are prone to producing unused, and thus under-trained, tokens due to the imbalance in repository and language diversity in the training data, as well as the dominance of source-specific, repetitive tokens that are often unusable in future inference. By modifying the BPE objective and introducing merge skipping, we implement different techniques under the name Source-Attributed BPE (SA-BPE) to regularize BPE training and minimize overfitting, thereby substantially reducing the number of under-trained tokens while maintaining the same inference procedure as with regular BPE. This provides an effective tool suitable for production use.

Modern language models are trained on large amounts of data. These data inevitably include controversial and stereotypical content, which contains all sorts of biases related to gender, origin, age, etc. As a result, the models express biased points of view or produce different results based on the assigned personality or the personality of the user. In this paper, we investigate various proxy measures of bias in large language models (LLMs). We find that evaluating models with pre-prompted personae on a multi-subject benchmark (MMLU) leads to negligible and mostly random differences in scores. However, if we reformulate the task and ask a model to grade the user’s answer, this shows more significant signs of bias. Finally, if we ask the model for salary negotiation advice, we see pronounced bias in the answers. With the recent trend for LLM assistant memory and personalization, these problems open up from a different angle: modern LLM users do not need to pre-prompt the description of their persona since the model already knows their socio-demographics.

Russenorsk, a pidgin language historically used in trade interactions between Russian and Norwegian speakers, represents a unique linguistic phenomenon. In this paper, we attempt to analyze its lexicon using modern large language models (LLMs), based on surviving literary sources. We construct a structured dictionary of the language, grouped by synonyms and word origins. Subsequently, we use this dictionary to formulate hypotheses about the core principles of word formation and grammatical structure in Russenorsk and show which hypotheses generated by large language models correspond to the hypotheses previously proposed ones in the academic literature. We also develop a “reconstruction” translation agent that generates hypothetical Russenorsk renderings of contemporary Russian and Norwegian texts.

Comics offer a compelling yet under-explored domain for computational narrative analysis, combining text and imagery in ways distinct from purely textual or audiovisual media. We introduce ComicScene154, a manually annotated dataset of scene-level narrative arcs derived from public-domain comic books spanning diverse genres. By conceptualizing comics as an abstraction for narrative-driven, multimodal data, we highlight their potential to inform broader research on multi-modal storytelling. To demonstrate the utility of ComicScene154, we present a baseline scene segmentation pipeline, providing an initial benchmark that future studies can build upon. Our results indicate that ComicScene154 constitutes a valuable resource for advancing computational methods in multimodal narrative understanding and expanding the scope of comic analysis within the Natural Language Processing community.

With the rise of computational social science, many scholars utilize data analysis and natural language processing tools to analyze social media, news articles, and other accessible data sources for examining political and social discourse. Particularly, the study of the emergence of echo-chambers due to the dissemination of specific information has become a topic of interest in mixed methods research areas. In this paper, we analyze data collected from two news portals, Breitbart News (BN) and New York Times (NYT) to prove the hypothesis that the formation of echo-chambers can be partially explained on the level of an individual information consumption rather than a collective topology of individuals’ social networks. Our research findings are presented through knowledge graphs, utilizing a dataset spanning 11.5 years gathered from BN and NYT media portals. We demonstrate that the application of knowledge representation techniques to the aforementioned news streams highlights, contrary to common assumptions, shows relative “internal” neutrality of both sources and polarizing attitude towards a small fraction of entities. Additionally, we argue that such characteristics in information sources lead to fundamental disparities in audience worldviews, potentially acting as a catalyst for the formation of echo-chambers.

This paper investigates the communication styles and structures of Twitter (X) communities within the vaccination context. While mainstream research primarily focuses on the echo-chamber phenomenon, wherein certain ideas are reinforced and participants are isolated from opposing opinions, this study reveals the presence of diverse communication styles across various communities. In addition to the communities exhibiting echo-chamber behavior, this research uncovers communities with distinct communication patterns. By shedding light on the nuanced nature of communication within social networks, this study emphasizes the significance of understanding the diversity of perspectives within online communities.

An empirical investigation into the simulation of the Big5 personality traits by large language models (LLMs), namely Llama-2, GPT-4, and Mixtral, is presented. We analyze the personality traits simulated by these models and their stability. This contributes to the broader understanding of the capabilities of LLMs to simulate personality traits and the respective implications for personalized human-computer interaction.

We show differences between a language-and-vision model CLIP and two text-only models — FastText and SBERT — when it comes to the encoding of individuation information. We study latent representations that CLIP provides for substrates, granular aggregates, and various numbers of objects. We demonstrate that CLIP embeddings capture quantitative differences in individuation better than models trained on text-only data. Moreover, the individuation hierarchy we deduce from the CLIP embeddings agrees with the hierarchies proposed in linguistics and cognitive science.

This paper evaluates the performance of Large Language Models (LLMs) in authorship attribu- tion and authorship verification tasks for Latin texts of the Patristic Era. The study showcases that LLMs can be robust in zero-shot author- ship verification even on short texts without sophisticated feature engineering. Yet, the mod- els can also be easily “mislead” by semantics. The experiments also demonstrate that steering the model’s authorship analysis and decision- making is challenging, unlike what is reported in the studies dealing with high-resource mod- ern languages. Although LLMs prove to be able to beat, under certain circumstances, the traditional baselines, obtaining a nuanced and truly explainable decision requires at best a lot of experimentation.

Evaluation plays a significant role in modern natural language processing. Most modern NLP benchmarks consist of arbitrary sets of tasks that neither guarantee any generalization potential for the model once applied outside the test set nor try to minimize the resource consumption needed for model evaluation. This paper presents a theoretical instrument and a practical algorithm to calculate similarity between benchmark tasks, we call this similarity measure “Vygotsky distance”. The core idea of this similarity measure is that it is based on relative performance of the “students” on a given task, rather that on the properties of the task itself. If two tasks are close to each other in terms of Vygotsky distance the models tend to have similar relative performance on them. Thus knowing Vygotsky distance between tasks one can significantly reduce the number of evaluation tasks while maintaining a high validation quality. Experiments on various benchmarks, including GLUE, SuperGLUE, CLUE, and RussianSuperGLUE, demonstrate that a vast majority of NLP benchmarks could be at least 40% smaller in terms of the tasks included. Most importantly, Vygotsky distance could also be used for the validation of new tasks thus increasing the generalization potential of the future NLP models.

Language models can greatly benefit from efficient tokenization. However, they still mostly utilize the classical Byte-Pair Encoding (BPE) algorithm, a simple and reliable method. BPE has been shown to cause such issues as under-trained tokens and sub-optimal compression that may affect the downstream performance. We introduce PickyBPE, a modified BPE algorithm that carries out vocabulary refinement during tokenizer training by removing merges that leave intermediate “junk” tokens. Our method improves vocabulary efficiency, eliminates under-trained tokens, and does not compromise text compression. Our experiments show that this method either improves downstream performance or does not harm it.

This study explores four methods of generating paraphrases in Malayalam, utilizing resources available for English paraphrasing and pre-trained Neural Machine Translation (NMT) models. We evaluate the resulting paraphrases using both automated metrics, such as BLEU, METEOR, and cosine similarity, as well as human annotation. Our findings suggest that automated evaluation measures may not be fully appropriate for Malayalam, as they do not consistently align with human judgment. This discrepancy underscores the need for more nuanced paraphrase evaluation approaches especially for highly agglutinative languages.

In this position paper, we contend that advancing our understanding of narrative and the effective generation of longer, subjectively engaging texts is crucial for progress in modern Natural Language Processing (NLP) and potentially the broader field of Artificial Intelligence. We highlight the current lack of appropriate datasets, evaluation methods, and operational concepts necessary for initiating work on narrative processing.

In the rapidly evolving landscape of Large Language Models (LLMs), introduction of well-defined and standardized evaluation methodologies remains a crucial challenge. This paper traces the historical trajectory of LLM evaluations, from the foundational questions posed by Alan Turing to the modern era of AI research. We categorize the evolution of LLMs into distinct periods, each characterized by its unique benchmarks and evaluation criteria. As LLMs increasingly mimic human-like behaviors, traditional evaluation proxies, such as the Turing test, have become less reliable. We emphasize the pressing need for a unified evaluation system, given the broader societal implications of these models. Through an analysis of common evaluation methodologies, we advocate for a qualitative shift in assessment approaches, underscoring the importance of standardization and objective criteria. This work serves as a call for the AI community to collaboratively address the challenges of LLM evaluation, ensuring their reliability, fairness, and societal benefit.

Current state-of-the-art NLP systems use large neural networks that require extensive computational resources for training. Inspired by human knowledge acquisition, researchers have proposed curriculum learning - sequencing tasks (task-based curricula) or ordering and sampling the datasets (data-based curricula) that facilitate training. This work investigates the benefits of data-based curriculum learning for large language models such as BERT and T5. We experiment with various curricula based on complexity measures and different sampling strategies. Extensive experiments on several NLP tasks show that curricula based on various complexity measures rarely have any benefits, while random sampling performs either as well or better than curricula.

The extensive surviving corpus of the ancient scholar Plutarch of Chaeronea (ca. 45-120 CE) also contains several texts which, according to current scholarly opinion, did not originate with him and are therefore attributed to an anonymous author Pseudo-Plutarch. These include, in particular, the work Placita Philosophorum (Quotations and Opinions of the Ancient Philosophers), which is extremely important for the history of ancient philosophy. Little is known about the identity of that anonymous author and its relation to other authors from the same period. This paper presents a BERT language model for Ancient Greek. The model discovers previously unknown statistical properties relevant to these literary, philosophical, and historical problems and can shed new light on this authorship question. In particular, the Placita Philosophorum, together with one of the other Pseudo-Plutarch texts, shows similarities with the texts written by authors from an Alexandrian context (2nd/3rd century CE).

This paper presents StoryDB — a broad multi-language dataset of narratives. StoryDB is a corpus of texts that includes stories in 42 different languages. Every language includes 500+ stories. Some of the languages include more than 20 000 stories. Every story is indexed across languages and labeled with tags such as a genre or a topic. The corpus shows rich topical and language variation and can serve as a resource for the study of the role of narrative in natural language processing across various languages including low resource ones. We also demonstrate how the dataset could be used to benchmark three modern multilanguage models, namely, mDistillBERT, mBERT, and XLM-RoBERTa.

This paper explores modern word embeddings in the context of sound symbolism. Using basic properties of the representations space one can construct semantic axes. A method is proposed to measure if the presence of individual sounds in a given word shifts its semantics of that word along a specific axis. It is shown that, in accordance with several experimental and statistical results, word embeddings capture symbolism for certain sounds.

This paper focuses on latent representations that could effectively decompose different aspects of textual information. Using a framework of style transfer for texts, we propose several empirical methods to assess information decomposition quality. We validate these methods with several state-of-the-art textual style transfer methods. Higher quality of information decomposition corresponds to higher performance in terms of bilingual evaluation understudy (BLEU) between output and human-written reformulations.

This paper shows that standard assessment methodology for style transfer has several significant problems. First, the standard metrics for style accuracy and semantics preservation vary significantly on different re-runs. Therefore one has to report error margins for the obtained results. Second, starting with certain values of bilingual evaluation understudy (BLEU) between input and output and accuracy of the sentiment transfer the optimization of these two standard metrics diverge from the intuitive goal of the style transfer task. Finally, due to the nature of the task itself, there is a specific dependence between these two metrics that could be easily manipulated. Under these circumstances, we suggest taking BLEU between input and human-written reformulations into consideration for benchmarks. We also propose three new architectures that outperform state of the art in terms of this metric.

This paper addresses author-stylized text generation. Using a version of a language model with extended phonetic and semantic embeddings for poetry generation we show that phonetics has comparable contribution to the overall model performance as the information on the target author. Phonetic information is shown to be important for English and Russian language. Humans tend to attribute machine generated texts to the target author.