EACL 2026 Papers

While chain-of-thought (CoT) prompting improves reasoning in large language models, its effectiveness in vision-language models (VLMs) remains limited due to over-reliance on textual cues and memorized knowledge. To investigate the visual reasoning capabilities of VLMs in complex real-world scenarios, we introduce DrivingVQA, a visual question answering dataset derived from driving theory exams, which contains 3,931 multiple-choice problems with expert-written explanations and grounded entities relevant to the reasoning process. Leveraging this dataset, we explore the benefits of incorporating entity-related information, such as entity names, spatial coordinates, and visual content, through supervised fine-tuning to enhance the model’s reasoning abilities. Our experiments demonstrate that interleaving textual explanations with visual tokens extracted from entities relevant to the question improves answer accuracy by 3.1% and reasoning accuracy by 4.6% over vanilla CoT prompting. Furthermore, we demonstrate that this retrieval-based approach effectively scales to the larger A-OKVQA reasoning dataset by leveraging automatically generated pseudo-labels, outperforming CoT prompting.

Emotion recognition in conversation (ERC) requires understanding both contextual dependencies and speaker-specific cues. Existing approaches often treat conversation context as a single representation or encode speaker identity shallowly, limiting their ability to capture fine-grained emotional dynamics. We propose PERC, a personality-aware ERC framework that (1) segregates conversational context into intra- and inter-speaker components, (2) models static or dynamic personality traits to represent stable and evolving speaker dispositions, and (3) performs contrastive cross-alignment between intra–intra and inter–inter representations to enforce contextual and personality consistency. Experiments on three ERC benchmarks show that PERC achieves new state-of-the-art performance, improving weighted F1 by up to 2.74% over non-LLM methods and 0.98% over recent LLM-based methods. Our results demonstrate the effectiveness of integrating context segregation, personality modeling, and contrastive alignment for emotion reasoning in dialogue.

The ability of LLM agents to plan and invoke tools exposes them to new safety risks, making a comprehensive red-teaming system crucial for discovering vulnerabilities and ensuring their safe deployment. We present SIRAJ, a generic red-teaming framework for arbitrary black-box LLM agents. We employ a dynamic two-step process that starts with an agent definition and generates diverse seed test cases that cover diverse risk outcomes, tool-use trajectories, and risk sources. Then, it iteratively constructs and refines model-based adversarial attacks based on the execution trajectories of former attempts. To optimize the red-teaming cost, we present a model distillation approach that leverages structured forms of a teacher model’s reasoning to train smaller models that are equally effective. Across diverse evaluation agent settings, our seed test case generation approach yields 2 – 2.5x boost to the coverage of risk outcomes and tool-calling trajectories. Our distilled 8B red-teamer model improves attack success rate by 100%, surpassing the 671B Deepseek-R1 model. Our ablations and analyses validate the effectiveness of the iterative framework, structured reasoning, and the generalization of our red-teamer models.

As language models evolve into autonomous agents that act and communicate on behalf of users, ensuring safety in multi-agent ecosystems becomes a central challenge. Interactions between personal assistants and external service providers expose a core tension between utility and protection: effective collaboration requires information sharing, yet every exchange creates new attack surfaces. We introduce ConVerse, a dynamic benchmark for evaluating privacy and security risks in agent–agent interactions. ConVerse spans three practical domains (travel, real estate, insurance) with 12 user personas and over 864 contextually grounded attacks (611 privacy, 253 security). Unlike prior single-agent settings, it models autonomous, multi-turn agent-to-agent conversations where malicious requests are embedded within plausible discourse. Privacy is tested through a three-tier taxonomy assessing abstraction quality, while security attacks target tool use and preference manipulation. Evaluating seven state-of-the-art models reveals persistent vulnerabilities—privacy attacks succeed in up to 88% of cases and security breaches in up to 60%—with stronger models leaking more. By unifying privacy and security within interactive multi-agent contexts, ConVerse reframes safety as an emergent property of communication.

Despite significant progress in natural image editing with state-of-the-art MLLMs, compositional layout and content editing for structured visual domains (e.g., posters, websites) remains underexplored. In this work, we introduce SMART-EDITOR, a multi-agent framework for compositional editing for structured images like posters or websites. Unlike prior models that focus on isolated local edits, SMART-EDITOR maintains global coherence through two complementary strategies: Reward-Refine, an inference-time reward-guided refinement method, and RewardDPO, a training-time preference optimization approach leveraging reward-aligned layout pairs. To evaluate performance, we introduce SMARTEdit-Bench, a benchmark of cascading multi-step edit instructions that are implicit in nature yet require layout and semantic-consistency preserving reasoning about edit order to preserve spatial and semantic consistency. Both automatic and human evaluations confirm the value of reward-guided planning in producing semantically consistent and visually coherent edits, beyond what single-shot VLMs can generate.

KV caching significantly improves the efficiency of Large Language Model (LLM) inference by storing attention states from previously processed tokens, enabling faster generation of subsequent tokens. However, as sequence length increases, the KV cache quickly becomes a major memory bottleneck. To address this, we propose PagedEviction, a novel fine-grained, structured KV cache pruning strategy that enhances the memory efficiency of vLLM’s PagedAttention. Unlike existing approaches that rely on attention-based token importance or evict tokens across different vLLM pages, PagedEviction introduces an efficient block-wise eviction algorithm tailored for paged memory layouts. Our method integrates seamlessly with PagedAttention without requiring any modifications to its CUDA attention kernels. We evaluate PagedEviction across Llama-3.1-8B-Instruct, Llama-3.2-1B-Instruct, and Llama-3.2-3B-Instruct models on the LongBench benchmark suite, demonstrating improved memory usage with better accuracy than baselines on long context tasks.

While large language models (LLMs) have achieved remarkable success in providing trustworthy responses for knowledge-intensive tasks, they still face critical limitations such as hallucinations and outdated knowledge. To address these issues, the retrieval-augmented generation (RAG) framework enhances LLMs with access to external knowledge via a retriever, enabling more accurate and real-time outputs about the latest events. However, this integration brings new security vulnerabilities: the risk that malicious content in the external database can be retrieved and used to manipulate model outputs. Although prior work has explored attacks on RAG systems, existing approaches either rely heavily on access to the retriever or fail to jointly consider both retrieval and generation stages, limiting their effectiveness, particularly in black-box scenarios. To overcome these limitations, we propose Token-level Precise Attack on the RAG (TPARAG), a novel framework that targets both white-box and black-box RAG systems. TPARAG leverages a lightweight white-box LLM as an attacker to generate and iteratively optimize malicious passages at the token level, ensuring both retrievability and high attack success in generation. Extensive experiments on open-domain QA datasets demonstrate that TPARAG consistently outperforms previous approaches in retrieval-stage and end-to-end attack effectiveness. These results further reveal critical vulnerabilities in RAG pipelines and offer new insights into improving their robustness.

Haptic captioning is the task of generating natural language descriptions from haptic signals, such as vibrations, for use in virtual reality and rehabilitation applications. While previous multimodal research has focused primarily on vision and audio, haptic feedback for the sense of touch remain underexplored. To address this gap, we formalize the haptic captioning task and propose HapticLLaMA, a multimodal sensory language model that interprets vibration signals into descriptions in a given sensory, emotional, or associative category. We investigate two types of haptic tokenizers, a frequency-based tokenizer and an EnCodec-based tokenizer, that convert haptic signals into sequences of discrete units, enabling their integration with the LLaMA model. HapticLLaMA is trained in two stages: (1) supervised fine-tuning using the LLaMA architecture with LoRA-based adaptation, and (2) fine-tuning via reinforcement learning from human feedback (RLHF). We assess HapticLLaMA’s captioning performance using both automated n-gram metrics and human evaluation.HapticLLaMA demonstrates strong capability in interpreting haptic vibration signals, achieving a METEOR score of 59.98 and a BLEU-4 score of 32.06, respectively. Furthermore, over 64% of the generated captions received human ratings above 3.5 on a 7-point scale, with RLHF yielding a 13% improvement in the overall rating distribution, indicating stronger alignment with human haptic perception. These findings highlight the potential of large language models to process and adapt to sensory data.

Automatic workflow generation is the process of automatically synthesizing sequences of LLM calls, tool invocations, and post-processing steps for complex end-to-end tasks. Most prior methods cast this task as an optimization problem with limited theoretical grounding. We propose to cast workflow generation as Bayesian inference over a posterior distribution on workflows, and introduce Bayesian Workflow Generation (BWG), a sampling framework that builds workflows step-by-step using parallel look-ahead rollouts for importance weighting and a sequential in-loop refiner for pool-wide improvements. We prove that, without the refiner, the weighted empirical distribution converges to the target posterior. We instantiate BWG as BayesFlow, a training-free algorithm for workflow construction. Across six benchmark datasets, BayesFlow improves accuracy by up to 9 percentage points over SOTA workflow generation baselines and by up to 65 percentage points over zero-shot prompting, establishing BWG as a principled upgrade to search-based workflow design.

Multi-task post-training of large language models (LLMs) is typically performed by mixing datasets from different tasks and optimizing them jointly. This approach implicitly assumes that all tasks contribute gradients of similar magnitudes; when this assumption fails, optimization becomes biased toward large-gradient tasks. In this paper, however, we show that this assumption fails in RL post-training: certain tasks produce significantly larger gradients, thus biasing updates toward those tasks. Such gradient imbalance would be justified only if larger gradients implied larger learning gains on the tasks (i.e., larger performance improvements)—but we find this is not true. Large-gradient tasks can achieve similar or even much lower learning gains than small-gradient ones. Further analyses reveal that these gradient imbalances cannot be explained by typical training statistics such as training rewards or advantages, suggesting that they arise from the inherent differences between tasks. This cautions against naive dataset mixing and calls for future work on principled gradient-level corrections for LLMs.

We introduce CLEAR-3K, a dataset of 3,008 assertion-reasoning questions designed to evaluate whether language models can determine if one statement causally explains another. Each question presents an assertion-reason pair and challenge language models to distinguish between semantic relatedness and genuine causal explanatory relationships. Through comprehensive evaluation of 21 state-of-the-art language models (ranging from 0.5B to 72B parameters), we identify two fundamental findings. First, language models frequently confuse semantic similarity with causality, relying on lexical and semantic overlap instead of inferring actual causal explanatory relationships. Second, as parameter size increases, models tend to shift from being overly skeptical about causal relationships to being excessively permissive in accepting them. Despite this shift, performance measured by the Matthews Correlation Coefficient plateaus at just 0.55, even for the best-performing models. Hence, CLEAR-3K provides a crucial benchmark for developing and evaluating causal explanatory reasoning in language models, which is an essential capability for applications that require accurate assessment of causal relationships.

Recent advances in audio generation led to an increasing number of deepfakes, making the general public more vulnerable to financial scams, identity theft, and misinformation. Audio deepfake detectors promise to alleviate this issue, with many recent studies reporting accuracy rates close to 99%. However, these methods are typically tested in an in-domain setup, where the deepfake samples from the training and test sets are produced by the same generative models. To this end, we introduce XMAD-Bench, a large-scale cross-domain multilingual audio deepfake benchmark comprising 668.8 hours of real and deepfake speech. In our novel dataset, the speakers, the generative methods, and the real audio sources are distinct across training and test splits. This leads to a challenging cross-domain evaluation setup, where audio deepfake detectors can be tested “in the wild”. Our in-domain and cross-domain experiments indicate a clear disparity between the in-domain performance of deepfake detectors, which is usually as high as 100%, and the cross-domain performance of the same models, which is sometimes similar to random chance. Our benchmark highlights the need for the development of robust audio deepfake detectors, which maintain their generalization capacity across different languages, speakers, generative methods, and data sources. Our benchmark is publicly released at https://github.com/ristea/xmad-bench/.

Instruction optimization provides a lightweight, model-agnostic approach to enhancing the reasoning performance of large language models (LLMs). This paper presents the first systematic comparison of instruction optimization, based on the DSPy optimization framework, for tabular fact verification. We evaluate four out-of-the-box prompting techniques that cover both text-only prompting and code use: direct prediction, Chain-of-Thought (CoT), ReAct with SQL tools, and CodeAct with Python execution. We study three optimizers from the DSPy framework—COPRO, MiPROv2, and SIMBA—across four benchmarks and three model families. We find that instruction optimization consistently improves verification accuracy, with MiPROv2 yielding the most stable gains for CoT, and SIMBA providing the largest benefits for ReAct agents, particularly at larger model scales. Behavioral analyses reveal that SIMBA encourages more direct reasoning paths by applying heuristics, thereby improving numerical comparison abilities in CoT reasoning and helping avoid unnecessary tool calls in ReAct agents. Across different prompting techniques, CoT remains effective for tabular fact checking, especially with smaller models. Although ReAct agents built with larger models can achieve competitive performance, they require careful instruction optimization.

Large Language Models (LLMs) excel at many reasoning tasks but struggle with knowledge-intensive queries due to their inability to dynamically access up-to-date or domain-specific information. Retrieval-Augmented Generation (RAG) has emerged as a promising solution, enabling LLMs to ground their responses in external sources. However, existing RAG methods lack fine-grained control over both the query and source sides, often resulting in noisy retrieval and shallow reasoning. In this work, we introduce DeepSieve, an agentic RAG framework that incorporates information sieving via LLM-as-a-knowledge-router. DeepSieve decomposes complex queries into structured sub-questions and recursively routes each to the most suitable knowledge source, filtering irrelevant information through a multi-stage distillation process. Our design emphasizes modularity, transparency, and adaptability, leveraging recent advances in agentic system design. Experiments on multi-hop QA tasks across heterogeneous sources demonstrate improved reasoning depth, retrieval precision, and interpretability over conventional RAG approaches.

Implicit Attribute Value Extraction (AVE) is essential for accurately representing products in e-commerce, as it infers lantent attributes from multimodal data. Despite advances in multimodal large language models (MLLMs), implicit AVE remains challenging due to the complexity of multidimensional data and gaps in vision-text understanding. In this work, we introduce MADIAVE, a multi-agent de- bate framework that employs multiple MLLM agents to iteratively refine inferences. Through a series of debate rounds, agents verify and up- date each other’s responses, thereby improving inference performance and robustness. Experi- ments on the ImplicitAVE dataset demonstrate that even a few rounds of debate significantly boost accuracy, especially for attributes with initially low performance. We systematically evaluate various debate configurations, includ- ing identical or different MLLM agents, and analyze how debate rounds affect convergence dynamics. Our findings highlight the poten- tial of multi-agent debate strategies to address the limitations of single-agent approaches and offer a scalable solution for implicit AVE in multimodal e-commerce.

Automatically extracting workflows as procedural graphs from natural language is a promising yet underexplored task that requires ensuring both structural validity and logical alignment. Recent advances in large language models (LLMs) show potential for graph extraction, but often yield ill-formed structures or misinterpret logical constructs such as gateways. We introduce , a multi-agent framework that treats procedural graph extraction as a multi-round reasoning process with structural and logical refinement agents. The framework operates in three iterative stages: (1) an LLM-based graph extraction phase, (2) a structural feedback phase where a simulation agent diagnoses and explains structural issues, and (3) a logical feedback phase where a semantic agent aligns semantics between flow logic and linguistic cues in the source text. Important feedback is prioritized and expressed in natural language, which is injected into the next-round prompt, enabling interpretable and controllable refinement. This modular design allows agents to target distinct error types without supervision or parameter updates. Experiments demonstrate that achieves substantial improvements in both structural correctness and logical consistency over strong baselines.

Information Retrieval (IR) is fundamental to many modern NLP applications. The rise of dense retrieval (DR), using neural networks to learn semantic vector representations, has significantly advanced IR performance. Central to training effective dense retrievers through contrastive learning is the selection of informative negative samples. Synthesizing 35 seminal papers, this survey provides a comprehensive and up-to-date overview of negative sampling techniques in dense IR. Our unique contribution is the focus on modern NLP applications and the inclusion of recent Large Language Model (LLM)-driven methods, an area absent in prior reviews. We propose a taxonomy that categorizes techniques, including random, static/dynamically mined, and synthetic datasets. We then analyze these approaches with respect to trade-offs between effectiveness, computational cost, and implementation difficulty. The survey concludes by outlining current challenges and promising future directions for the use of LLM-generated synthetic data.

Large Language Models (LLMs) have enabled remarkable progress in natural language processing, yet their high computational and memory demands pose challenges for deployment in resource-constrained environments. Although recent low-rank decomposition methods offer a promising path for structural compression, they often suffer from accuracy degradation, expensive calibration procedures, and result in inefficient model architectures that hinder real-world inference speedups. In this paper, we propose FLAT-LLM, a fast and accurate, training-free structural compression method based on fine-grained low-rank transformations in the activation space. Specifically, we reduce the hidden dimension by transforming the weights using truncated eigenvectors computed via head-wise Principal Component Analysis, and employ a greedy budget redistribution strategy to adaptively allocate ranks across decoders. FLAT-LLM achieves efficient and effective weight compression without recovery fine-tuning, which could complete the calibration within a few minutes.Evaluated across 5 models and 11 datasets, FLAT-LLM outperforms structural pruning baselines in generalization and downstream performance, while delivering inference speedups over decomposition-based methods.

Large Language Models (LLMs) are increasingly used to answer factual, information-seeking questions (ISQs). While prior work often focuses on false, misleading information, little attention has been paid to true but strategically persuasive content that can derail a model’s reasoning. To address this gap, we introduce a new evaluation dataset, TruthTrap, in two languages, i.e., English and Farsi, on Iran-related ISQs, each paired with a correct explanation and a persuasive-yet-misleading true hint. We then evaluate nine diverse LLMs (spanning proprietary and open-source systems) via factuality classification and multiple-choice QA tasks, finding that accuracy drops by 25%, on average, when models encounter these misleading yet factual hints. Also, the models’ predictions match the hint-aligned options up to 77 percent of the time. Notably, models often misjudge such hints in isolation yet still integrate them into final answers. Our results highlight a significant limitation in LLM outputs, underscoring the importance of robust fact-verification and emphasizing real-world risks posed by partial truths in domains like social media, education, and policy-making.