EACL 2026 Papers

A key barrier to interpreting large language models is polysemanticity, where neurons activate for multiple unrelated concepts. Sparse autoencoders (SAEs) have been proposed to mitigate this issue by transforming dense activations into sparse, more interpretable features. While prior work suggests that SAEs promote monosemanticity, no quantitative comparison has examined how concept activation distributions differ between SAEs and their base models. This paper provides the first systematic evaluation of SAEs against base models through activation distribution lens. We introduce a fine-grained concept separability score based on the Jensen–Shannon distance, which captures how distinctly a neuron’s activation distributions vary across concepts. Using two large language models (Gemma-2-2B and DeepSeek-R1) and multiple SAE variants across five datasets (including word-level and sentence-level), we show that SAEs reduce polysemanticity and achieve higher concept separability. To assess practical utility, we evaluate concept-level interventions using two strategies: full neuron masking and partial suppression. We find that, compared to base models, SAEs enable more precise concept-level control when using partial suppression. Building on this, we propose Attenuation via Posterior Probabilities (APP), a new intervention method that uses concept-conditioned activation distributions for targeted suppression. APP achieves the smallest perplexity increase while remaining highly effective at concept removal.

Generative retrieval (GR) differs from the traditional index–then–retrieve pipeline by storing relevance in model parameters and generating retrieval cues directly from the query, but it can be brittle out of domain and expensive to scale. We introduce QueStER (QUEry SpecificaTion for gEnerative Keyword-Based Retrieval), which bridges GR and query reformulation by learning to generate explicit keyword-based search specifications. Given a user query, a lightweight LLM produces a keyword query that is executed by a standard retriever (BM25), combining the generalization benefits of generative query rewriting with the efficiency and scalability of lexical indexing. We train the rewriting policy with reinforcement learning techniques. Across in- and out-of-domain evaluations, QueStER consistently improves over BM25 and is competitive with neural IR baselines, while maintaining strong efficiency.

Vision–language models (VLMs) have demonstrated strong reasoning abilities in literal multimodal tasks such as visual mathematics and science question answering. However, figurative language—such as sarcasm, humor, and metaphor—remains a significant challenge, as it conveys intent and emotion through subtle incongruities between expressed and intended meanings. In multimodal settings, accompanying images can amplify or invert textual meaning, demanding models that reason across modalities and account for subjectivity.We propose a three-step framework for developing efficient multimodal reasoning models that can (i) interpret multimodal figurative language, (ii) provide transparent reasoning traces, and (iii) generalize across multiple figurative styles. Experiments across four styles show that (1) incorporating reasoning traces substantially improves multimodal figurative understanding, (2) reasoning learned in one style can transfer to others—especially between related styles like sarcasm and humor, and (3) training jointly across styles yields a generalized reasoning VLM that outperforms much larger open- and closed-source models.Our findings show that lightweight VLMs with verifiable reasoning achieve robust cross-style generalization while providing inspectable reasoning traces for multimodal tasks. The code and implementation are available at https://github.com/scheshmi/CrossStyle-MMR.

The discourse around toxicity and LLMs in NLP largely revolves around detection tasks. This work shifts the focus to evaluating LLMs’ reasoning about toxicity—from their explanations that justify a stance—to enhance their trustworthiness in downstream tasks. Despite extensive research on explainability, it is not straightforward to adopt existing methods to evaluate free-form toxicity explanation due to their over-reliance on input text perturbations, among other challenges. To account for these, we propose a novel, theoretically-grounded multi-dimensional criterion, Argument-based Consistency (ArC), that measures the extent to which LLMs’ free-form toxicity explanations reflect an ideal and logical argumentation process. Based on uncertainty quantification, we develop six metrics for ArC to comprehensively evaluate the (in)consistencies in LLMs’ toxicity explanations. We conduct several experiments on three Llama models (of size up to 70B) and an 8B Ministral model on five diverse toxicity datasets. Our results show that while LLMs generate plausible explanations to simple prompts, their reasoning about toxicity breaks down when prompted about the nuanced relations between the complete set of reasons, the individual reasons, and their toxicity stances, resulting in inconsistent and irrelevant responses. We open-source our code and LLM-generated explanations for future works.

Complex claim verification requires decomposing sentences into verifiable subclaims, yet existing methods struggle to align decomposition quality with verification performance. We propose a reinforcement learning (RL) approach that jointly optimizes decomposition quality and verifier alignment using Group Relative Policy Optimization (GRPO). Our method integrates: (i) structured sequential reasoning; (ii) supervised finetuning on teacher-distilled exemplars; and (iii) a multi-objective reward balancing format compliance, verifier alignment, and decomposition quality. Across six evaluation settings, our trained 8B decomposer improves downstream verification performance to 71.75% macro-F1, outperforming prompt-based approaches (+1.99, +6.24) and existing RL methods (+5.84). Human evaluation confirms the high quality of the generated subclaims. Our framework enables smaller language models to achieve state-of-the-art claim verification by jointly optimising for verification accuracy and decomposition.

We present FeatEng, a novel benchmark designed to evaluate the ability of large language models (LLMs) to perform feature engineering, a critical and knowledge-intensive task in data science. FeatEng assesses LLMs by their capacity to generate Python code that transforms raw tabular data into features that improve the performance of a downstream machine learning model. Our analysis of LLM outputs reveals that success on FeatEng often requires the application of significant world and domain knowledge, along with complex reasoning, to construct novel data representations. While focused on feature engineering, the benchmark probes a confluence of abilities indicative of an LLM’s broader potential for practical, data-centric problem-solving. We demonstrate that FeatEng offers a targeted and efficient approach to assess a specific but crucial aspect of LLM capabilities relevant to real-world data science applications.

One of the first pre-processing steps for constructing web-scale LLM pretraining datasets involves extracting text from HTML. Despite the immense diversity of web content, existing open-source datasets predominantly apply a single fixed extractor to all webpages. In this work, we investigate whether this practice leads to suboptimal coverage and utilization of Internet data. We first show that while different extractors may lead to similar model performance on standard language understanding tasks, the pages surviving a fixed filtering pipeline can differ substantially. This suggests a simple intervention: by taking a Union over different extractors, we can increase the token yield of DCLM-Baseline by up to 71% while maintaining benchmark performance. We further show that for structured content such as tables and code blocks, extractor choice can significantly impact downstream task performance, with differences of up to 10 percentage points (p.p.) on WikiTQ and 3 p.p. on HumanEval.

Self-attention mechanisms in transformers enable tokens to interact across a sequence but lack an explicit inductive bias to capture local contextual dependencies, an inherent characteristic of natural languages. We propose Token-Wise Kernels (TWiKers), a novel enhancement to transformers that learn token-specific convolutional kernels applied to the keys or values. Each token is assigned a small kernel, initialized to the “Central Dirac” (e.g., [0,1,0] for size=3), meaning the token “bears” the attention from all other tokens alone. During training, these kernels adapt, and greater deviation from the Central Dirac indicates stronger attention redistribution to neighboring tokens. This introduces the first transformer weights with direct semantic interpretability. Our experiments show that content words (e.g., nouns and verbs) retain self-focus, while function words (e.g., prepositions and conjunctions) shift attention toward their neighbors, aligning with their syntactic and semantic roles. We further apply TWiKers to distinguish literary genres, historical periods, and authors, demonstrating their effectiveness in capturing high-level stylistic patterns. Finally, we demonstrate the potential of TWiKers as an effective inductive bias to improve transformer training, validated across a range of downstream tasks.

The rapid integration of large language models (LLMs) into high-stakes legal work has exposed a critical gap: no benchmark exists to systematically stress-test their reliability against the nuanced, adversarial, and often subtle flaws present in real-world contracts. To address this, we introduce CLAUSE, a first-of-its-kind benchmark designed to evaluate the fragility of an LLM’s legal reasoning. We study the capabilities of LLMs to detect and reason about fine-grained discrepancies by producing over 7500 real-world perturbed contracts from foundational datasets like CUAD and ContractNLI. Our novel, persona-driven pipeline generates 10 distinct anomaly categories, which are then validated against official statutes using a Retrieval-Augmented Generation (RAG) system to ensure legal fidelity. We use CLAUSE to evaluate leading LLMs’ ability to detect embedded legal flaws and explain their significance. Our analysis shows a key weakness: these models often miss subtle errors and struggle even more to justify them legally. Our work outlines a path to identify and correct such reasoning failures in legal AI.

Multimodal reasoning tasks such as visual question answering (VQA) require models to process both language and visual inputs. However, existing approaches typically decompose only language queries, treating images as monolithic inputs. We introduce REDI, a framework that jointly decomposes both images and questions into visual sub-domains (segmentation, material, depth, and color) with corresponding sub-questions. REDI uses an MLLM orchestrator to select the sub-domains required for each query, generate domain-specific sub-questions with grounded object references (via shared object labels), and fuse worker outputs via consistency-aware aggregation (verify–refine–override) to produce the final answer. This hierarchical multi-agent design mitigates error propagation and improves compositional reasoning across both open- and closed-source MLLMs. On SEEDBench, MMBench, and CLEVR, REDI achieves absolute accuracy improvements of 8.9%, 8.2%, and 16.0% over chain-of-thought and visual programming baselines. Project webpage: https://madhav-kanda.github.io/redi

We demonstrate that embeddings derived from large language models, when processed with “Survey and Questionnaire Item Embeddings Differentials” (SQuID), can recover the structure of human values obtained from human rater judgments on the Revised Portrait Value Questionnaire (PVQ-RR). We compare multiple embedding models across a number of evaluation metrics including internal consistency, dimension correlations and multidimensional scaling configurations. Unlike previous approaches, SQuID addresses the challenge of obtaining negative correlations between dimensions without requiring domain-specific fine-tuning or training data re-annotation. Quantitative analysis reveals that our embedding-based approach explains 55% of variance in dimension-dimension similarities compared to human data. Multidimensional scaling configurations show alignment with pooled human data from 49 different countries. Generalizability tests across three personality inventories (IPIP, BFI-2, HEXACO) demonstrate that SQuID consistently increases correlation ranges, suggesting applicability beyond value theory. These results show that semantic embeddings can effectively replicate psychometric structures previously established through extensive human surveys. The approach offers substantial advantages in cost, scalability and flexibility while maintaining comparable quality to traditional methods. Our findings have significant implications for psychometrics and social science research, providing a complementary methodology that could expand the scope of human behavior and experience represented in measurement tools.

Large Language Models (LLMs) are now state-of-the-art at summarization, yet the internal notion of importance that drives their information selections remains hidden. We propose to investigate this by combining behavioral and computational analyses. Behaviorally, we generate a series of length-controlled summaries for each document and derive empirical importance distributions based on how often each information unit is selected. These reveal that LLMs converge on consistent importance patterns, sharply different from pre-LLM baselines, and that LLMs cluster more by family than by size. Computationally, we identify that certain attention heads align well with empirical importance distributions, and that middle-to-late layers are strongly predictive of importance. Together, these results provide initial insights into what LLMs prioritize in summarization and how this priority is internally represented, opening a path toward interpreting and ultimately controlling information selection in these models.

Information Extraction (IE), encompassing Named Entity Recognition (NER), Named Entity Linking (NEL), and Relation Extraction (RE), is critical for transforming the rapidly growing volume of scientific publications into structured, actionable knowledge. This need is especially evident in fast-evolving biomedical fields such as the gut-brain axis, where research investigates complex interactions between the gut microbiota and brain-related disorders. Existing biomedical IE benchmarks, however, are often narrow in scope and rely heavily on distantly supervised or automatically generated annotations, limiting their utility for advancing robust IE methods. We introduce GutBrainIE, a benchmark based on more than 1,600 PubMed abstracts, manually annotated by biomedical and terminological experts with fine-grained entities, concept-level links, and relations. While grounded in the gut-brain axis, the benchmark’s rich schema, multiple tasks, and combination of highly curated and weakly supervised data make it broadly applicable to the development and evaluation of biomedical IE systems across domains.

Data visualizations like charts are fundamental tools for quantitative analysis and decision-making across fields, requiring accurate interpretation and mathematical reasoning. The emergence of Multimodal Large Language Models (MLLMs) offers promising capabilities for automated visual data analysis, such as processing charts, answering questions, and generating summaries. However, they provide no visibility into which parts of the visual data informed their conclusions; this black-box nature poses significant challenges to real-world trust and adoption. In this paper, we take the first major step toward evaluating and enhancing the capabilities of MLLMs to attribute their reasoning process by highlighting the specific regions in charts and graphs that justify model answers. To this end, we contribute RADAR, a semi-automatic approach to obtain a benchmark dataset comprising 1000 charts, 2000 question-answer pairs, 3599 reasoning steps, and 11,220 attribution annotations. We also introduce a method that provides attribution for chart-based mathematical reasoning. Experimental results demonstrate that our reasoning-guided approach improves attribution accuracy by up to 15 percentage points compared to baseline methods, and enhanced attribution capabilities translate to stronger answer generation, achieving high semantic similarity (BERTScore \sim 0.90) with ground truth responses. This advancement represents a significant step toward more interpretable and trustworthy chart analysis systems, enabling users to verify and understand model decisions through reasoning and attribution.

Retrieval-Augmented Generation (RAG) has emerged as a standard framework for knowledge-intensive NLP tasks, combining large language models (LLMs) with document retrieval from external corpora. Despite its widespread use, most RAG pipelines continue to treat retrieval and reasoning as isolated components, retrieving documents once and then generating answers without further interaction. This static design often limits performance on complex tasks that require iterative evidence gathering or high-precision retrieval. Recent work in both the information retrieval (IR) and NLP communities has begun to close this gap by introducing adaptive retrieval and ranking methods that incorporate feedback. In this survey, we present a structured overview of advanced retrieval and ranking mechanisms that integrate such feedback. We categorize feedback signals based on their source and role in improving the query, retrieved context, or document pool. By consolidating these developments, we aim to bridge IR and NLP perspectives and highlight retrieval as a dynamic, learnable component of end-to-end RAG systems.

In this paper we explore where information is collected and how it is propagated throughout layers in large language models (LLMs). We begin by examining the surprising computational importance of punctuation tokens which previous work has identified as attention sinks and memory aids. Using intervention-based techniques, we evaluate the necessity and sufficiency of punctuation tokens across layers in GPT-2, DeepSeek, and Gemma. Our results show stark model-specific differences: for GPT-2, punctuation is both necessary and sufficient in multiple layers, while this holds far less in DeepSeek and not at all in Gemma. Extending beyond punctuation, we ask whether LLMs process different components of input (e.g., subjects, adjectives, punctuation, full sentences) by forming early static summaries reused across the network, or if the model remains sensitive to changes in these components across layers. We investigate whether different reasoning rules are processed differently by LLMs. In particular, through interchange intervention and layer-swapping experiments, we find that conditional statements (if, then), and universal quantification (for all) are processed very differently. Our findings offer new insight into the internal mechanisms of punctuation usage and reasoning in LLMs and have implications for interpretability and model analysis.

Large Language Models (LLMs) frequently generate hallucinated content, posing significant challenges for applications where factuality is crucial. While existing hallucination detection methods typically operate at the sentence level or passage level, we propose FactSelfCheck, a novel zero-resource black-box sampling-based method that enables fine-grained fact-level detection. Our approach represents text as interpretable knowledge graphs consisting of facts in the form of triples, providing clearer insights into content factuality than traditional approaches. Through analyzing factual consistency across multiple LLM responses, we compute fine-grained hallucination scores without requiring external resources or training data. Our evaluation demonstrates that FactSelfCheck performs competitively with leading sentence-level sampling-based methods while providing more detailed and interpretable insights. Most notably, our fact-level approach significantly improves hallucination correction, achieving a 35.5% increase in factual content compared to the baseline, while sentence-level SelfCheckGPT yields only a 10.6% improvement. The granular nature of our detection enables more precise identification and correction of hallucinated content. Additionally, we contribute FavaMultiSamples, a novel dataset that addresses a gap in the field by providing the research community with a second dataset for evaluating sampling-based methods.

The intersection of AI and legal systems presents a growing need for tools that support legal education, particularly in under-resourced languages such as Romanian. In this work, we aim to evaluate the capabilities of Large Language Models (LLMs) and Vision-Language Models (VLMs) in understanding and reasoning about the Romanian driving law through textual and visual question-answering tasks. To facilitate this, we introduce RoD-TAL, a novel multimodal dataset comprising Romanian driving test questions, text-based and image-based, along with annotated legal references and explanations written by human experts. We implement and assess retrieval-augmented generation (RAG) pipelines, dense retrievers, and reasoning-optimized models across tasks, including Information Retrieval (IR), Question Answering (QA), Visual IR, and Visual QA. Our experiments demonstrate that domain-specific fine-tuning significantly enhances retrieval performance. At the same time, chain-of-thought prompting and specialized reasoning models improve QA accuracy, surpassing the minimum passing grades required for driving exams. We highlight the potential and limitations of applying LLMs and VLMs to legal education. We release the code and resources through the GitHub repository (https://github.com/vladman-25/RoD-TAL).