EACL 2026 Papers

Few forces rival fear in their ability to mobilize societies, distort communication, and reshape collective behavior. In computational linguistics, fear is primarily studied as an emotion, but not as a distinct form of speech. Fear speech content is widespread and growing, and often outperforms hate-speech content in reach and engagement because it appears “civiler” and evades moderation. Yet the computational study of fear speech remains fragmented and under-resourced. This can be understood by recognizing that fear speech is a phenomenon shaped by contributions from multiple disciplines. In this paper, we bridge cross-disciplinary perspectives by comparing theories of fear from Psychology, Political science, Communication science, and Linguistics. Building on this, we review existing definitions. We follow up with a survey of datasets from related research areas and propose a taxonomy that consolidates different dimensions of fear for studying fear speech. By reviewing current datasets and defining core concepts, our work offers both theoretical and practical guidance for creating datasets and advancing fear speech research.

Extending large language models to low-resource languages is essential for global accessibility, but training separate models per language is prohibitively expensive. Mixture-of-Experts (MoE) architectures address this by adding sparse language-specific parameters, but determining how many experts each layer needs remains an open question. Current approaches allocate experts based on layer-level similarity, yet language processing exhibits fine-grained specialization at individual neurons. We propose \textbf{NeuronMoE}, a method that analyzes language-specific neurons across all transformer components to guide expert allocation per layer based on empirically measured cross-lingual neuron diversity. Applied to Llama-3.2-3B for low-resource languages (Greek, Turkish, and Hungarian), this approach achieves approximately 40% average parameter reduction while matching the performance of the LayerMoE baseline. We find that low-resource language experts independently develop neuron specialization patterns mirroring the high-resource language, which are concentrated in early and late layers. This reveals potential universal architectural principles in how multilingual models organize linguistic knowledge. Our approach generalizes across architectures, as validated on Qwen, and shows that allocation strategy matters more than total expert count.

Large Language Models (LLMs) need to be in accordance with human values—being helpful, harmless, and honest (HHH)—is important for safe deployment. Existing works use Supervised Fine-Tuning (SFT) and Mixture-of-Experts (MoE) to align LLMs. However, these works face challenges in multi-objective settings, such as SFT leading to interference between conflicting objectives, while MoEs suffer from miscalibrated routing. We term this failure mode Axis Collapse, marked by(1) disjoint feature spaces causing catastrophic forgetting, and (2) unreliable inference from misrouted experts. To resolve this, we propose AlignX, a two-stage framework. Stage 1 uses prompt-injected fine-tuning to extract axis-specific task features, mitigating catastrophic forgetting. Stage 2 deploys a MoCaE module that calibrates expert routing using fractal and natural geometry, improving inference reliability. AlignX achieves significant gains on Alpaca (Helpfulness), BeaverTails (Harmlessness), and TruthfulQA (Honesty), with +171.5% win rate, +110.1% in truthfulness-informativeness, and 4.3% fewer safety violations. It also reduces latency and memory usage by over 35% compared to prior MoEs. Results across four LLMs validate its generalizability. Code and data are available at: https://github.com/gskgautam/AlignX

Chat assistants increasingly integrate web search functionality, enabling them to retrieve and cite external sources. While this promises more reliable answers, it also raises the risk of amplifying misinformation from low-credibility sources. In this paper, we introduce a novel methodology for evaluating assistants’ web search behavior, focusing on source credibility and the groundedness of responses with respect to cited sources. Using 100 claims across five misinformation-prone topics, we assess GPT-4o, GPT-5, Perplexity, and Qwen Chat. Our findings reveal differences between the assistants, with Perplexity achieving the highest source credibility, whereas GPT-4o exhibits elevated citation of non-credible sources on sensitive topics. This work provides the first systematic comparison of commonly used chat assistants for fact-checking behavior, offering a foundation for evaluating AI systems in high-stakes information environments.

Knowledge Graph-based Retrieval-Augmented Generation (KG-RAG) is an increasingly explored approach for combining the reasoning capabilities of large language models with the structured evidence of knowledge graphs. However, current evaluation practices fall short: existing benchmarks often include questions that can be directly answered using existing triples in KG, making it unclear whether models perform reasoning or simply retrieve answers directly. Moreover, inconsistent evaluation metrics and lenient answer matching criteria further obscure meaningful comparisons. In this work, we introduce a general method for constructing benchmarks and present BRINK (Benchmark for Reasoning under Incomplete Knowledge) to systematically assess KG-RAG methods under knowledge incompleteness. Our empirical results show that current KG-RAG methods have limited reasoning ability under missing knowledge, often rely on internal memorization, and exhibit varying degrees of generalization depending on their design.

Intrinsic evaluation metrics for conditional language models, such as perplexity or bits-per-character, are widely used in both mono- and multilingual settings. These metrics are rather straightforward to use and compare in monolingual setups, but rest on a number of assumptions in multilingual setups. One such assumption is that comparing the perplexity of CLMs on parallel sentences is indicative of their quality since the information content (here understood as the semantic meaning) is the same. However, the metrics are inherently measuring information content in the information-theoretic sense. We make this and other such assumptions explicit and discuss their implications. We perform experiments with six metrics on two multi-parallel corpora both with mono- and multilingual models. Ultimately, we find that current metrics are not universally comparable. We look at the form-meaning debate to provide some explanation for this.

Explicit representations of predicate-argument relations form the basis of interpretable semantic analysis, supporting reasoning, generation, and evaluation. However, attaining such semantic structures requires costly annotation efforts and has remained largely confined to English. We leverage the Question-Answer driven Semantic Role Labeling (QA-SRL) framework — a natural-language formulation of predicate-argument relations — as the foundation for extending semantic annotation to new languages. To this end, we introduce a cross-linguistic projection approach that reuses an English QA-SRL parser within a constrained translation and word-alignment pipeline to automatically generate question-answer annotations aligned with target-language predicates. Applied to Hebrew, Russian, and French — spanning diverse language families — the method yields high-quality training data and fine-tuned, language-specific parsers that outperform strong multilingual LLM baselines (GPT-4o, LLaMA-Maverick). By leveraging QA-SRL as a transferable natural-language interface for semantics, our approach enables efficient and broadly accessible predicate-argument parsing across languages.

Medical reasoning in large language models seeks to replicate clinicians’ cognitive processes in interpreting patient data and making diagnostic decisions. However, widely used benchmarks—such as MedQA, MedMCQA, and PubMedQA—mix questions that require multi-step reasoning with those answerable through factual recall, complicating evaluation. We introduce an expert-validated evaluation framework that disentangles knowledge recall from reasoning by training a PubMedBERT-based classifier and applying it to 11 widely used biomedical QA benchmarks. This framework reveals that only 32.8% of questions require multi-step reasoning, indicating that current evaluations largely measure factual recall. Stratified evaluation of biomedical models (HuatuoGPT-o1, MedReason, m1) and general-domain models (DeepSeek-R1, o4-mini, Qwen3) consistently shows lower performance on reasoning-heavy than knowledge-heavy questions (e.g., HuatuoGPT-o1: 56.9% on knowledge vs.44.8% on reasoning). Beyond aggregate accuracy, we assess robustness through adversarial evaluations in which models are prefixed with uncertainty-inducing, incorrect statements; biomedical reasoning models degrade sharply in this setting (e.g., MedReason: 50.4% to 24.4%), with declines especially pronounced on reasoning-heavy questions. Finally, we show that fine-tuning on high-quality, reasoning-heavy examples augmented with adversarial traces, followed by reinforcement learning with GRPO, improves both robustness and accuracy across knowledge and reasoning subsets within our evaluation framework.

There are many methods and systems to tackle the ontology alignment problem, yet a major challenge persists in producing high-quality mappings among a set of input ontologies. Adopting a human-in-the-loop approach during the alignment process has become essential in applications requiring very accurate mappings. However, user involvement is expensive when dealing with large ontologies. In this paper, we analyse the feasibility of using Large Language Models (LLM) to aid the ontology alignment problem. LLMs are used only in the validation of a subset of correspondences for which there is high uncertainty. We have conducted an extensive analysis over several tasks of the Ontology Alignment Evaluation Initiative (OAEI), reporting in this paper the performance of several state-of-the-art LLMs using different prompt templates. Using LLMs as resulted in strong performance in the OAEI 2025, achieving the top-2 overall rank in the bio-ml track.

Despite major advances in multilingual modeling, large quality disparities persist across languages. Besides the obvious impact of uneven training resources, typological properties have also been proposed to determine the intrinsic difficulty of modeling a language. The existing evidence, however, is mostly based on small monolingual language models or bilingual translation models trained from scratch. We expand on this line of work by analyzing two large pre-trained multilingual translation models, NLLB-200 and Tower+, which are state-of-the-art representatives of encoder-decoder and decoder-only machine translation, respectively. Based on a broad set of languages, we find that target language typology drives translation quality of both models, even after controlling for more trivial factors, such as data resourcedness and writing script. Additionally, languages with certain typological properties benefit more from a wider search of the output space, suggesting that such languages could profit from alternative decoding strategies beyond the standard left-to-right beam search. To facilitate further research in this area, we release a set of fine-grained typological properties for 212 languages of the FLORES+ MT evaluation benchmark.

We investigate the automated evaluation of English language learner answers to writing tasks featuring picture stories.This is usually limited to language proficiency only, neglecting the context of the picture. Instead, our analysis focuses on task adherence, which for example allows detection of off-topic answers.Since there is a lack of suitable training and evaluation data, our first step is to build the PictureStories dataset.To this end, we develop a marking rubric that covers task adherence with respect to both form and content. Six annotators mark 713 learner answers written in response to one of five picture stories.Having assembled the dataset, we then explore to what extent task adherence can be predicted automatically. Our experiments assume a scenario where no or just a few labelled answers are available for the picture story which is being marked.For form-focused criteria, we find that it is beneficial to finetune models across tasks.With content-focused criteria, few-shot prompting Qwen emerges as the best-performing method. We examine the trade-off between including the story image vs. example answers in the prompt and find that examples suffice in many cases. While for some LLMs, few-shot prompting results may look promising on the surface, we demonstrate that a much simpler method can do just as well when shown the same examples.

The way our eyes move while reading can tell us about the cognitive effort required to process the text. In the present study, we use this fact to generate texts with controllable reading ease. Our method employs a model that predicts human gaze patterns to steer language model outputs towards eliciting certain reading behaviors. We evaluate the approach in an eye-tracking experiment with native and non-native speakers of English. The results demonstrate that the method is effective at making the generated texts easier or harder to read, measured both in terms of reading times and perceived difficulty of the texts. A statistical analysis reveals that the changes in reading behavior are mostly due to features that affect lexical processing. Possible applications of our approach include text simplification for information accessibility and generation of personalized educational material for language learning.

Reliable uncertainty quantification (UQ) is essential when employing large language models (LLMs) in high-risk domains such as clinical question answering (QA). In this work, we evaluate uncertainty estimation methods for clinical QA focusing, for the first time, on eleven clinical specialties and six question types, and across ten open-source LLMs (general-purpose, biomedical, and reasoning models), alongside representative proprietary models. We analyze score-based UQ methods, present a case study introducing a novel lightweight method based on behavioral features derived from reasoning-oriented models, and examine conformal prediction as a complementary set-based approach. Our findings reveal that uncertainty reliability is not a monolithic property, but one that depends on clinical specialty and question type due to shifts in calibration and discrimination. Our results highlight the need to select or ensemble models based on their distinct, complementary strengths and clinical use.

Health wellness agents typically rely on large language models (LLMs) for response generation, where contextual information from a user’s conversation history can be used for response grounding and personalization. High-quality conversation summaries are one such method which can reduce the number of input tokens during response generation, decreasing overhead and inference latency. However, directly purposing LLMs for this task is infeasible due to the scale of the task, the compute overhead, and health data compliance regulations. Furthermore, ground truth for real-world datasets is scarce due to privacy concerns and the high cost of health expert annotators. These factors necessitate the development of small, potentially on-device, language models capable of health dialogue summarization, particularly in the absence of ground truth labels. In this paper, we first present a comprehensive empirical study that benchmarks a variety of state-of-the-art smaller language models to better understand their baseline capabilities. Second, we present an unsupervised method that uses the summaries from multiple models, refined with iterative feedback, to generate high-quality summaries of health dialogues. Experiments illustrate that our method is able to outperform baseline on both open-source and proprietary benchmarks. Notably, our method can be run viably on local compute without a GPU, using just a single Macbook with 16 GB of memory.

Mixture-of-Experts (MoE) based large language models (LLMs) have gained popularity due to their multi-task capability, where each input token activates only a subset of “expert” subnetworks. However, whether each expert can truly specialize to a certain task remains poorly understood, while activation analysis shows frequent cross-layer co-activation of experts for the same input, resembling a collaborative behavior. In this paper, we use a dictionary learning approach to show that experts in MoE LLMs form hierarchical and semantically coherent collaborative groups that correspond to specific linguistic and cognitive functions (e.g., mathematical reasoning, syntactic processing), mirroring specialized functional region observed in neuroscience. Furthermore, leveraging these discovered expert groups enables significant model compression with minimal performance degradation, outperforming existing methods by 2.5% while enabling up to 50% expert reduction. These findings provide the first systematic analysis of expert collaboration mechanisms in MoE LLMs, revealing that specialization emerges from joint activation of experts across all layers. We further developed an interactive visualization platform that enables researchers to explore expert collaboration patterns and their semantic associations.

The Korean legal system mandates public access to court judgments to ensure judicial transparency. However, this requirement conflicts with privacy protection obligations due to the prevalence of Personally Identifiable Information (PII) in legal documents. To address this challenge, we introduce K-LegalDeID, a large-scale benchmark dataset and an efficient KLUEBERT-CRF model for de-identification for Korean court judgments. Our primary contribution is a new large-scale benchmark dataset spanning 39 legal domains, with its quality is validated by a high inter-annotator agreement (IAA) with Fleiss’ Kappa of 0.7352. Our results demonstrate that a lightweight KLUEBERT-CRF model, when trained on our dataset, achieves state-of-the-art performance with an entity-level micro F1 score of 0.9923. Our end-to-end framework offers a practical and computationally efficient solution for real-world legal systems.

We present AraBabyTalk-EGY, an enriched release of the Egyptian Arabic CHILDES corpus, that opens the child-adult interactions genre to modern Arabic NLP research. Starting from the original CHILDES recordings and IPA transcriptions of caregiver-child sessions, we (i) map each IPA token to fully diacritized Arabic script, and (ii) add core part-of-speech tags and lemmas aligned with existing dialectal Arabic morphological resources. These layers yield ~26K annotated tokens suitable for both text- and speech-based NLP tasks. We provide a benchmark on morphological disambiguation and Arabic ASR. We outline lexical and morphosyntactic differences between AraBabyTalk-EGY and general Egyptian Arabic resources, highlighting the value of genre-specific training data for language acquisition studies and Arabic speech technology.

Despite recent advances, long-sequence video generation frameworks still suffer from significant limitations: poor assistive capability, suboptimal visual quality, and limited expressiveness. To mitigate these limitations, we propose MAViS, an end-to-end multi-agent collaborative framework for long-sequence video storytelling. MAViS orchestrates specialized agents across multiple stages, including script writing, shot designing, character modeling, keyframe generation, video animation, and audio generation. In each stage, agents operate under the 3E Principle—Explore, Examine, and Enhance—to ensure the completeness of intermediate outputs. Considering the capability limitations of current generative models, we propose the Script Writing Guidelines to optimize compatibility between scripts and generative tools. Experimental results demonstrate that MAViS achieves state-of-the-art performance in assistive capability, visual quality, and video expressiveness. Its modular framework further enables scalability with diverse generative models and tools. With just a brief prompt, MAViS enables users to rapidly explore diverse visual storytelling and creative directions for sequential video generation by efficiently producing high-quality, complete long-sequence videos. To the best of our knowledge, MAViS is the only framework that provides multimodal design output – videos with narratives and background music.

Large language model (LLM) systems increasingly power everyday AI applications such as chatbots, computer-use assistants, and autonomous robots, where performance often depends on manually well-crafted prompts. LLM-based prompt optimizers reduce that effort by iteratively refining prompts from scored feedback, yet the security of this optimization stage remains underexamined. We present the first systematic analysis of poisoning risks in LLM-based prompt optimization. Using HarmBench, we find systems are substantially more vulnerable to manipulated feedback than to query poisoning alone: feedback-based attacks raise attack success rate (ASR) by up to ΔASR = 0.48. We introduce a simple fake reward attack that requires no access to the reward model and significantly increases vulnerability. We also propose a lightweight highlighting defense that reduces the fake reward ΔASR from 0.23 to 0.07 without degrading utility. These results establish prompt optimization pipelines as a first-class attack surface and motivate stronger safeguards for feedback channels and optimization frameworks.