EACL 2026 论文列表

Generalized Category Discovery (GCD) is a practical and challenging open-world task that aims to recognize both known and novel categories in unlabeled data using limited labeled data from known categories. Due to the lack of supervision, previous GCD methods face significant challenges, such as difficulty in rectifying errors for confusing instances, and inability to effectively uncover and leverage the semantic meanings of discovered clusters. Therefore, additional annotations are usually required for real-world applicability. However, human annotation is extremely costly and inefficient. To address these issues, we propose DeLFGCD , a unified framework for generalized category discovery that actively learns from diverse and collaborative LLM feedback. Our approach leverages three different types of LLM feedback to: (1) improve instance-level contrastive features, (2) generate category descriptions, and (3) align uncertain instances with LLM-selected category descriptions. Extensive experiments demonstrate the superior performance of DeLFGCD over state-of-the-art models across diverse datasets, metrics, and supervision settings.

Extending Large Language Models (LLMs) to new languages is challenging, with most methods proposed suffering from high computational cost and catastrophic forgetting of original model capabilities. Embedding relearning (CITATION), a technique that creates new tokenizers and tunes embeddings on fixed model weights for target language adaptation, is both light-weight and performant. However, it has only been shown to work for older generation encoder-only models and for high resource languages. In this paper, we extend this framework to decoder-only LLMs focusing on joint adaptation to many languages, including low-resource ones. We experiment in three language groups over 100 languages each. We adapt a pre-trained LLM via switching to a customized tokenizer, and relearning the embedding layer. Across 96 diverse languages spanning both classification and generation tasks, we show embedding relearning improves models by up to 20%, being highly competitive with full-weight updating baselines while vastly more computationally efficient and mitigating catastrophic forgetting. This translates into better results in transferring the improved multilingual performance to tasks that build on core English abilities (e.g., multilingual math reasoning), compared to various baselines. Further analysis reveals the critical role of customizing tokenizers in achieving effective language transfer, particularly for non-Latin script languages.

Large-scale Vision-Language Models (LVLMs) are being deployed in real-world settings that require visual inference. As capabilities improve, applications in navigation, education, and accessibility are becoming practical. These settings require accommodation of perceptual variation rather than assuming a uniform visual experience. Color perception illustrates this requirement: it is central to visual understanding yet varies across individuals due to Color Vision Deficiencies, an aspect largely ignored in multimodal AI.In this work, we examine whether LVLMs can account for variation in color perception using the Ishihara Test. We evaluate model behavior through generation, confidence, and internal representation, using Ishihara plates as controlled stimuli that expose perceptual differences. Although models possess factual knowledge about color vision deficiencies and can describe the test, they fail to reproduce the perceptual outcomes experienced by affected individuals and instead default to normative color perception. These results indicate that current systems lack mechanisms for representing alternative perceptual experiences, raising concerns for accessibility and inclusive deployment in multimodal settings.

In-context learning depends not only on what appears in the prompt but also on when it appears. To isolate this temporal component from semantic confounds, we construct prompts with repeated anchor tokens and average the model’s predictions over hundreds of random permutations of the intervening context. This approach ensures that any observed position-dependent effects are driven purely by temporal structure rather than token identity or local semantics. Across four transformer LLMs and three state-space/recurrent models, we observe a robust serial recall signature: models allocate disproportionate probability mass to the tokens that previously followed the anchor, but the strength of this signal is modulated by serial position, yielding model-specific primacy/recency profiles. We then introduce an overlapping-episode probe in which only a short cue from one episode is re-presented; retrieval is reliably weakest for episodes embedded in the middle of the prompt, consistent with “lost-in-the-middle” behavior. Mechanistically, ablating high-induction-score attention heads in transformers reduces serial recall and episodic separation. For state-space models, ablating a small fraction of high-attribution channels produces analogous degradations, suggesting a sparse subspace supporting induction-style copying. Together, these results clarify how temporal biases shape retrieval across architectures and provide controlled probes for studying long-context behavior.

Entity structure extraction, which aims to extract entities and their associated attribute–value structures from text, is an essential task for text understanding and knowledge graph construction. Existing methods based on large language models (LLMs) typically rely heavily on predefined entity attribute schemas or annotated datasets, often leading to incomplete extraction results. To address these challenges, we introduce ZOES, a novel approach to entity structure extraction that does not require any schema or annotated samples. ZOES operates via a principled mechanism of enrichment, refinement, and unification, based on the insight that an entity and its associated structure are mutually reinforcing. Experiments demonstrate that ZOES consistently enhances LLMs’ ability to extract more complete entity structures across three different domains, showcasing both the effectiveness and generalizability of the method. These findings suggest that such an enrichment, refinement, and unification mechanism may serve as a principled approach to improving the quality of LLM-based entity structure discovery in various scenarios.

Large language model (LLM) agents are increasingly deployed to tackle complex tasks, often necessitating collaboration among multiple specialized agents. However, multi-agent collaboration introduces new challenges in planning, coordination, and verification. Execution failures frequently arise not from flawed reasoning alone, but from subtle misalignments in task interpretation, output format, or inter-agent handoffs. To address these challenges, we present VeriMAP, a framework for multi-agent collaboration with verification-aware planning. The VeriMAP planner decomposes tasks, models subtask dependencies, and encodes planner-defined passing criteria as subtask verification functions (VFs) in Python and natural language. We evaluate VeriMAP on diverse datasets, demonstrating that it outperforms both single- and multi-agent baselines while enhancing system robustness and interpretability. Our analysis highlights how verification-aware planning enables reliable coordination and iterative refinement in multi-agent systems, without relying on external labels or annotations.

Diet plays a central role in human health, and Nutrition Question Answering (QA) offers a promising path toward personalized dietary guidance and the prevention of diet-related chronic diseases. However, existing methods face two fundamental challenges: the limited reasoning capacity of single-agent systems and the complexity of designing effective multi-agent architectures, as well as contextual overload that hinders accurate decision-making. We introduce Nutritional-Graph Router (NG-Router), a novel framework that formulates nutritional QA as a supervised, knowledge-graph–guided multi-agent collaboration problem. NG-Router integrates agent nodes into heterogeneous knowledge graphs and employs a graph neural network to learn task-aware routing distributions over agents, leveraging soft supervision derived from empirical agent performance. To further address contextual overload, we propose a gradient-based subgraph retrieval mechanism that identifies salient evidence during training, thereby enhancing multi-hop and relational reasoning. Extensive experiments across multiple benchmarks and backbone models demonstrate that NG-Router consistently outperforms both single-agent and ensemble baselines, offering a principled approach to domain-aware multi-agent reasoning for complex nutritional health tasks.

Large language model (LLM) routing has emerged as a crucial strategy for balancing computational costs with performance by dynamically assigning queries to the most appropriate model based on query complexity. Despite recent advances showing that preference-data-based routers can outperform traditional methods, current evaluation benchmarks remain limited—they largely focus on general model capabilities while overlooking task-specific behaviors and critical concerns such as privacy, safety, and potential backdoor vulnerabilities introduced through preference data. In response, we propose the DSC benchmark: Diverse, simple, and categorized, an evaluation framework that categorizes router performance across a broad spectrum of query types—including coding, translation, mathematics, human instructions, general knowledge, and LLM jailbreaking—and integrates privacy and safety assessments to reveal hidden risks. Our experiments on three preference-based routers and two commercial counterparts demonstrate that while these systems improve efficiency, they often make suboptimal, category-driven decisions; for instance, a BERT-based router directs all coding and mathematics queries to the most powerful LLM—even when simpler models would suffice—while routing jailbreaking attempts to weaker models, thereby elevating safety risks.

Psychological insights have long shaped pivotal NLP breakthroughs, from attention mechanisms to reinforcement learning and social modeling. As Large Language Models (LLMs) develop, there is a rising consensus that psychology is essential for capturing human-like cognition, behavior, and interaction.This paper reviews how psychological theories can inform and enhance stages of LLM development. Our review integrates insights from six subfields of psychology, including cognitive, developmental, behavioral, social, personality psychology, and psycholinguistics. With stage-wise analysis, we highlight current trends and gaps in how psychological theories are applied. By examining both cross-domain connections and points of tension, we aim to bridge disciplinary divides and promote more thoughtful integration of psychology into NLP research.

While multimodal large language models (LLMs) excel at dialogue, whether they can adequately parse the structure of conversation—conversational roles and threading—remains underexplored. In this work, we introduce a suite of tasks and release TV-MMPC, a new annotated dataset, for multimodal conversation structure understanding. Our evaluation reveals that while all multimodal LLMs outperform our heuristic baseline, even the best-performing model we consider experiences a substantial drop in performance when character identities of the conversation are anonymized. Beyond evaluation, we carry out a sociolinguistic analysis of 350,842 utterances in TVQA. We find that while female characters initiate conversations at rates in proportion to their speaking time, they are 1.2 times more likely than men to be cast as an addressee or side-participant, and the presence of side-participants shifts the conversational register from personal to social.

Personalization in LLMs often relies on costly human feedback or interaction logs, limiting scalability and neglecting deeper user attributes. We introduce (Generative Response with Aligned Values, Interests, and Traits of You), a framework for generating synthetic, profile-grounded preference data that captures users’ interests, values, beliefs, and personality traits. By integrating demographic, cultural, and psychological frameworks—including Hofstede’s cultural dimensions, Schwartz’s basic values, the World Values Survey, and Big Five OCEAN traits—synthesizes chosen/rejected preference pairs to guide personalized content generation. We evaluate on book descriptions for 400 Amazon users, comparing it to prompt-based conditioning, standard fine-tuning, and naive synthetic pair generation. Profile-grounded synthetic data consistently improves generation, especially across multiple cultures (USA, Brazil, Japan, India), achieving over 4% higher preference gains across baselines, with user studies showing that outputs are preferred over 86% of the time. Our results show that scenario-grounded synthetic data can capture richer user variation, reduce reliance on costly annotation, and produce more engaging, user-centered content, offering a scalable path for LLM personalization. Code and datasets will be released upon acceptance.

Language models (LMs) excel at tasks across diverse domains, yet require substantial computational resources during inference. Compression techniques such as pruning and quantization offer a practical path towards efficient LM deployment, exemplified by their ability to preserve performance on general-purpose benchmarks. However, general-purpose LM compression methods can negatively affect performance in specialized domains (e.g. biomedical or legal). Recent work has sought to address this issue, but requires a computationally expensive full-parameter fine-tuning pipeline. To this end, we propose MixCal, a novel calibration method designed to improve the in-domain performance of compressed LMs in a post-training setting. Through extensive experimentation, we demonstrate that MixCal substantially outperforms existing approaches on domain-specific tasks while preserving general performance. Notably, these performance gains are achieved while also reducing the computational cost of LM compression.

A wide array of NLP/NLU models have been developed for the Persian language and have shown promising results. However, the performance of such models drops significantly when applied to the colloquial form of Persian. This challenge arises from the substantial differences between colloquial and formal Persian and the lack of parallel data facilitating the robustness of the model to the colloquial data or to transform the data to formal Persian. In addressing this gap, our research is dedicated to the development of the HarfoSokhan dataset, a large-scale colloquial to formal Persian parallel dataset of 6M sentence pairs. Our proposed dataset is a critical resource for training models that can effectively bridge the linguistic variations between colloquial and formal Persian. To illustrate the utility of our dataset, we used it to train a GPT2 model, which exhibited remarkable proficiency in colloquial to formal text style transfer, outperforming both OpenAI’s GPT-3.5-turbo model and a leading rule-based system in this task. This conclusion is supported by our proposed ranking-based human evaluation. The results underscore the significance of the HarfoSokhan dataset in enhancing the performance of natural language processing models in the challenging task of colloquial to formal Persian conversion.

Large Language Models (LLMs) are increasingly embedded in the daily lives of individuals across diverse social classes. This widespread integration raises urgent concerns about the subtle, implicit biases these models may contain. In this work, we investigate such biases through the lens of ethical reasoning, analyzing model responses to scenarios in a new dataset we propose comprising 1,016 scenarios, systematically categorized into ethical, unethical, and neutral types. Our study focuses on dimensions that are socially influential but less explored, including (i) residency status, (ii) political ideology, (iii) Fitness Status, (iv) educational attainment, and (v) attitudes toward AI. To assess LLMs’ behavior, we propose a baseline and employ one statistical test and one metric: a permutation test that reveals the presence of bias by comparing the probability distributions of ethical/unethical scenarios with the probability distribution of neutral scenarios on each demographic group, and a tendency measurement that captures the magnitude of bias with respect to the relative difference between probability distribution of ethical and unethical scenarios. Our evaluations of 12 prominent LLMs reveal persistent and nuanced biases across all four attributes, and Llama models exhibited the most pronounced biases. These findings highlight the need for refined ethical benchmarks and bias-mitigation tools in LLMs.

Large Language Models (LLMs) increasingly mediate our social, cultural, and political interactions. While they can simulate some aspects of human behavior and decision-making, it is still underexplored whether repeated interactions with other agents amplify their biases or lead to exclusionary behaviors. To this end, we study Chirper.ai—an LLM-driven social media platform—analyzing 7M posts and interactions among 32K LLM agents over a year. We start with homophily and social influence among LLMs, learning that similar to humans’, their social networks exhibit these fundamental phenomena. Next, we study the toxic language of LLMs, its linguistic features, and their interaction patterns, finding that LLMs show different structural patterns in toxic posting than humans. After studying the ideological leaning in LLMs posts, and the polarization in their community, we focus on how to prevent their potential harmful activities. We present a simple yet effective method, called Chain of Social Thought (CoST), that reminds LLM agents to avoid harmful posting.

Multimodal Large Language Models (MLLMs) show reasoning promise, yet their visual perception is a critical bottleneck. Paradoxically, MLLMs sometimes produce correct answers while misinterpreting crucial visual elements, masking these underlying perception failures. Our preliminary analysis on a joint perception-reasoning dataset revealed that 29% of correct reasoning answers from a leading MLLM contained perception errors. To systematically study visual perception abilities of MLLMs, we introduce Do You See Me- a scalable, programmatically generated benchmark with 1758 images and 2612 questions across seven core subtasks spanning 2D and 3D variants (twelve total tasks) providing parametric control over difficulty levels. The benchmark tasks are inspired by human psychology. Our evaluation of eleven leading MLLMs reveals a stark deficit: humans achieve 95.83% accuracy, while top MLLMs average below 50%. This performance gap widens drastically as task complexity increases. Further diagnostics show: (1) supervised finetuning offers only modest gains (11%), (2) models tend to exploit task “shortcuts” like MCQ formats over detailed visual analysis, and (3) Chain-of-Thought prompting can degrade complex visual tasks by verbalizing images into lossy text. These findings expose the foundational perception limits in current MLLMs and highlight the need for robust visual perception improvements in MLLMs. The benchmark dataset, source code and evaluation scripts are available at[https://github.com/microsoft/Do-You-See-Me].

Machine translation (MT) systems universally degrade when faced with code-mixed text. This problem is more acute for low-resource languages that lack dedicated parallel corpora. This work directly addresses this gap for Vietnamese-English, a language context characterized by challenges including orthographic ambiguity and the frequent omission of diacritics in informal text. We introduce VietMix, the first expert-translated, naturally occurring parallel corpus of Vietnamese-English code-mixed text. We establish VietMix’s utility by developing a data augmentation pipeline that leverages iterative fine-tuning and targeted filtering. Experiments show that models augmented with our data outperform strong back-translation baselines by up to +3.5 xCOMET points and improve zero-shot models by up to +11.9 points. Our work delivers a foundational resource for a challenging language pair and provides a validated, transferable framework for building and augmenting corpora in other low-resource settings.

We present a comprehensive evaluation of large language models’ (LLMs) ability to process culturally grounded language, specifically to understand and pragmatically use figurative expressions that encode local knowledge and social nuance. Using figurative language as a proxy for cultural nuance and local knowledge, we design evaluation tasks for contextual understanding, pragmatic use, and connotation interpretation across Arabic and English. We evaluate 22 open- and closed-source LLMs on Egyptian Arabic idioms, multidialectal Arabic proverbs, and English proverbs. Results show a consistent hierarchy: accuracy on Arabic proverbs is 4.29% lower than on English proverbs, and performance on Egyptian idioms is 10.28% lower than on Arabic proverbs. On the pragmatic use task, accuracy drops by 14.07% relative to understanding, though providing idioms’ contextual sentences improves accuracy by 10.66%. Models also struggle with connotative meaning, reaching at most 85.58% agreement with human annotators on idioms with full inter-annotator agreement. Figurative language thus serves as an effective diagnostic for cultural reasoning, revealing that while LLMs often interpret figurative meaning, they still face major challenges in using it appropriately. To support future research, we release Kinayat, the first dataset of Egyptian Arabic idioms designed for both figurative understanding and pragmatic use evaluation.

Attribution and fact verification are critical challenges in natural language processing for assessing information reliability. While automated systems and Large Language Models (LLMs) aim to retrieve and select concise evidence to support or refute claims, they often present users with either insufficient or overly redundant information, leading to inefficient and error-prone verification. To address this, we propose Evidence Ranking, a novel task that prioritizes presenting sufficient information as early as possible in a ranked list. This minimizes user reading effort while still making all available evidence accessible for sequential verification. We compare two approaches for the new ranking task: one-shot ranking and incremental ranking. We introduce a new evaluation framework, inspired by information retrieval metrics, and construct a unified benchmark by aggregating existing fact verification datasets. Extensive experiments with diverse models show that incremental ranking strategies better capture complementary evidence and that LLM-based methods outperform shallower baselines, while still facing challenges in balancing sufficiency and redundancy. Compared to evidence selection, we conduct a controlled user study and demonstrate that evidence ranking both reduces reading effort and improves verification. This work provides a foundational step toward more interpretable, efficient, and user-aligned information verification systems.