EACL 2026 Papers

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

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

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

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

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

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

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

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

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

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

Since many real-world documents combine textual and tabular data, robust Retrieval Augmented Generation (RAG) systems are essential for effectively accessing and analyzing such content to support complex reasoning tasks. Therefore, this paper introduces T^2-RAGBench, a benchmark comprising \textbf{23,088} question-context-answer triples, designed to evaluate RAG methods on real-world text-and-table data. Unlike typical QA datasets that operate under \textit{Oracle Context} settings, T^2-RAGBench challenges models to first retrieve the correct context before conducting numerical reasoning. Existing QA datasets containing text-and-table data typically contain context-dependent questions, which may yield multiple correct answers depending on the provided context. To address this, we transform SOTA datasets into a context-independent format, validated by experts as 91.3% context-independent questions, enabling reliable RAG evaluation. Our comprehensive evaluation identifies \textit{Hybrid BM25} , a technique that combines dense and sparse vectors, as the most effective approach for text-and-table data. However, results demonstrate that T^2-RAGBench remains challenging even for SOTA LLMs and RAG methods. Further ablation studies examine the impact of embedding models and corpus size on retrieval performance. T^2-RAGBench provides a realistic and rigorous benchmark for existing RAG methods on text-and-table data. Code and dataset are available online: https://github.com/uhh-hcds/g4kmu-paper

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

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

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

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

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

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

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