EACL 2026 论文列表

Predicting how audiences react to Arabic social media posts requires reasoning beyond textual sentiment: reactions emerge from collective interpretation moderated by engagement dynamics and topical context. We present a multi-task learning (MTL) framework that jointly learns (i) audience reaction classification (Love, Haha, Angry, Sad, Care, Wow), (ii) engagement magnitude regression (six reactions, comments, shares), and (iii) non-engagement detection. On a corpus of 158k Arabic Facebook posts spanning women’s rights, gender debates, and economic empowerment, our model achieves a test macro-F1 of 72.4 and weighted-F1 of 89.1.

Multilingual Fine-tuning of Large Language Models (LLMs) has achieved great advancements in machine translation. However, existing research focuses only on the traditional fine-tuning setting with a fixed set of languages, lacking dynamic adaptability to new ones. Introducing new languages requires retraining and often causes catastrophic forgetting. In this study, we propose a completely modular fine-tuning pipeline that enables dynamic language adaptation for LLMs. Instead of directly fine-tuning on all languages, our approach first trains English-centric input and output LoRA adapters for each language separately, and then merges the corresponding adapters for arbitrary-direction translation without any additional training. Experiments on 12 translation directions of four low-resource and less-supported languages show that modular fine-tuning achieves up to 86% performance of traditional multi-parallel full-parameter fine-tuning, while training only 0.1% parameters and relying solely on English-centric data without any catastrophic forgetting. Furthermore, we perform a comprehensive analysis about the merging ratio, when to merge, and the rationale for using English as a bridge language via Bayesian Optimization and logit lens.

Large Language Model (LLM) hallucinations are usually treated as defects of the model or its decoding strategy. Drawing on classical linguistics, we argue that a query’s form can also shape a listener’s (and model’s) response. We operationalize this insight by constructing a 22-dimension query feature vector covering clause complexity, lexical rarity, and anaphora, negation, answerability, and intention grounding, all known to affect human comprehension. Using 369,837 real-world queries, we ask: Are there certain types of queries that make hallucination more likely? A large-scale analysis reveals a consistent “risk landscape”: certain features such as deep clause nesting and underspecification align with higher hallucination propensity. In contrast, clear intention grounding and answerability align with lower hallucination rates. Others, including domain specificity, show mixed, dataset- and model-dependent effects. Thus, these findings establish an empirically observable query-feature representation correlated with hallucination risk, paving the way for guided query rewriting and future intervention studies.

Negation is a fundamental linguistic phenomenon that poses ongoing challenges for Large Language Models (LLMs), particularly in tasks requiring deep semantic understanding. Current benchmarks often treat negation as a minor detail within broader tasks, such as natural language inference. Consequently, there is a lack of benchmarks specifically designed to evaluate comprehension of negation. In this work, we introduce Thunder-NUBench — a novel benchmark explicitly created to assess sentence-level understanding of negation in LLMs. Thunder-NUBench goes beyond identifying surface-level cues by contrasting standard negation with structurally diverse alternatives, such as local negation, contradiction, and paraphrase. This benchmark includes manually created sentence-negation pairs and a multiple-choice dataset, allowing for a comprehensive evaluation of models’ understanding of negation.

Speculative decoding has emerged as an effective method to reduce latency and inference cost of LLM inferences. However, there has been inadequate attention towards the energy requirements of these models. To address this gap, this paper presents a comprehensive survey of energy requirements of speculative decoding strategies, with detailed analysis on how various factors – model size and family, speculative decoding strategies, and dataset characteristics – influence the energy optimizations.

Recent advancements in Large Language Models (LLMs) have enabled autonomous agents to decompose complex tasks, select appropriate tools, and execute structured workflows. However, a key challenge in this field is the lack of a universal, large-scale, and cross-domain benchmark to systematically evaluate LLMs’ ability to reason over and utilize interconnected tools for automation. Existing benchmarks, such as TaskBench, focus on manually curated tool graphs for benchmark generation, which lack scalability and diversity across domains. To address this, we propose UniToolBench, a benchmark that incorporates automated tool graph construction by formulating link prediction as a probabilistic task, instead of relying on categorical LLM outputs. Furthermore, we introduce a confidence-based beam search sampling strategy to select high-confidence tool dependencies, ensuring more structured and semantically coherent subgraphs for evaluation. Through extensive experiments on multiple datasets, we demonstrate that while LLMs show promise in tool selection, significant challenges remain in parameter prediction and handling complex tool dependencies.

Large Language Models (LLMs) have demonstrated remarkable capabilities, enabling language agents to excel at single-turn tasks. However, their application to complex, multi-step, and long-horizon tasks remains challenging. While reinforcement learning (RL) offers a promising avenue for addressing these challenges, mainstream approaches typically rely solely on sparse, outcome-based rewards — a limitation that becomes especially problematic for group-based RL algorithms lacking critic models, such as Group Relative Policy Optimization (GRPO). In such methods, uniformly rewarding or penalizing all actions within a trajectory can lead to training instability and suboptimal policies, because beneficial and detrimental actions are often entangled across multi-step interactions. To address this challenge, we propose SALT, a novel and lightweight framework that provides a finer-grained advantage assignment, derived solely from outcome rewards. We achieve this by constructing a graph from trajectories of the same prompt, which allows us to quantify the quality of each step and assign advantages accordingly. Crucially, SALT is designed as a plug-and-play module that seamlessly integrates with existing group-based RL algorithms — requiring no modifications to the rollout procedure and introducing negligible computational overhead. Extensive experiments on the WebShop, ALFWorld, and AppWorld benchmarks with various model sizes demonstrate that SALT consistently improves performance. We also conduct a thorough analysis to validate the design choices behind SALT and offer actionable insights.

Recent advancements in Large Vision-Language Models (LVLMs) have shown groundbreaking capabilities across diverse multimodal tasks. However, these models remain vulnerable to adversarial jailbreak attacks, where adversaries craft subtle perturbations to bypass safety mechanisms and trigger harmful outputs. Existing white-box attacks methods require full model accessibility, suffer from computing costs and exhibit insufficient adversarial transferability, making them impractical for real-world, black-box settings. To address these limitations, we propose a black-box jailbreak attack on LVLMs via Zeroth-Order optimization using Simultaneous Perturbation Stochastic Approximation (ZO-SPSA). ZO-SPSA provides three key advantages: (i) gradient-free approximation by input-output interactions without requiring model knowledge, (ii) model-agnostic optimization without the surrogate model and (iii) lower resource requirements with reduced GPU memory consumption. We evaluate ZO-SPSA on three LVLMs, including InstructBLIP, LLaVA and MiniGPT-4, achieving the highest jailbreak success rate of 83.0% on InstructBLIP, while maintaining imperceptible perturbations comparable to white-box methods. Moreover, adversarial examples generated from MiniGPT-4 exhibit strong transferability to other LVLMs, with ASR reaching 64.18%. These findings underscore the real-world feasibility of black-box jailbreaks and expose critical weaknesses in the safety mechanisms of current LVLMs.

Neural audio codecs have recently enabled high-fidelity reconstruction at high compression rates, especially for speech. However, speech and non-speech audio exhibit fundamentally different spectral characteristics: speech energy concentrates in narrow bands around pitch harmonics (80-400 Hz), while non-speech audio requires faithful reproduction across the full spectrum, particularly preserving higher frequencies that define timbre and texture. This poses a challenge—speech-optimized neural codecs suffer degradation on music or sound. Treating the full spectrum holistically is suboptimal: frequency bands have vastly different information density and perceptual importance by content type, yet full-band approaches apply uniform capacity across frequencies without accounting for these acoustic structures. To address this gap, we propose BSCodec (Band-Split Codec), a novel neural audio codec architecture that splits the spectral dimension into separate bands and compresses each band independently. Experimental results demonstrate that BSCodec achieves superior reconstruction over baselines across sound and music, while maintaining competitive quality in the speech domain, when trained on the same combined dataset of speech, music and sound. Downstream benchmark tasks further confirm that BSCodec shows strong potential for use in downstream applications.

Text-to-Image models may generate harmful content, such as pornographic images, particularly when unsafe prompts are submitted. To address this issue, safety filters are often added on top of text-to-image models, or the models themselves are aligned to reduce harmful outputs. However, these defenses remain vulnerable when an attacker strategically designs adversarial prompts to bypass these safety guardrails. In this work, we propose PromptTune, a method to jailbreak text-to-image models with safety guardrails using a fine-tuned large language model. Unlike other query-based jailbreak attacks that require repeated queries to the target model, our attack generates adversarial prompts efficiently after fine-tuning our AttackLLM. We evaluate our method on three datasets of unsafe prompts and against five safety guardrails. Our results demonstrate that our approach effectively bypasses safety guardrails, outperforms existing no-box attacks, and also facilitates other query-based attacks. Our code is available at https://github.com/zhengyuan-jiang/PromptTune.

The ability to generate SPARQL queries from natural language questions is crucial for ensuring efficient and accurate retrieval of structured data from knowledge graphs (KG). While large language models (LLMs) have been widely adopted for SPARQL query generation, they are often susceptible to hallucinations and out-of-distribution errors when generating KG elements, such as Uniform Resource Identifiers (URIs), based on opaque internal parametric knowledge. We propose PGMR (Post-Generation Memory Retrieval), a modular framework where the LLM produces an intermediate query using natural language placeholders for URIs, and a non-parametric memory module is subsequently employed to retrieve and resolve the correct KG URIs. PGMR significantly enhances query correctness (SQM) across various LLMs, datasets, and distribution shifts, while achieving the near-complete suppression of URI hallucinations. Critically, we demonstrate PGMR’s superior safety and robustness: a retrieval confidence threshold enables PGMR to effectively refuse to answer queries that lack support, and the retriever proves highly resilient to memory noise, maintaining strong performance even when the non-parametric memory size is scaled up to 9 times with irrelevant, distracting entities.

Code-switching, where speakers alternate between languages within a single utterance, poses unique challenges for language identification (LID). Existing LID models often fail to reliably identify English spoken with the accent of the matrix (dominant) language. We show that finetuning LID models with small amounts of such accented English significantly improves code-switched LID, without degrading performance on standard monolingual speech—a limitation observed with direct finetuning on code-switched utterances. This is achieved via low-rank adaptation (LoRA) on limited accented data, which allows models to adapt efficiently. To better evaluate performance, we introduce LangRank, a metric that captures the relative ranking of identified languages often overlooked by traditional metrics. Our method generalizes across multiple language pairs, including Hindi-English, Bengali-English, Mandarin-English, and Arabic-English, providing robust LID in code-switched multilingual contexts.

Large language models (LLMs) are increasingly used as knowledge discovery tools. Humanistic disciplines like historical linguistics and literary studies have shown interest in this capability. These fields often construct arguments on the basis of distinctions between phenomena like time-period or genre. Such methodological investments complicate reliance on LLMs pretrained over large sets of broadly-collected data. We show that efficient pretraining techniques produce useful models of semantic change over modest historical corpora without allowing potential contamination from anachronistic data. We verify that these trained-from-scratch models better respect historical divisions and are more computationally efficient compared to the standard approach of fine-tuning an existing LLM. We compare the trade-offs in general linguistic fluency versus detecting and characterizing various forms of linguistic change, and provide a pipeline implementation of our approach that can be readily adapted and applied to a wide range of diachronic phenomena.

With the widespread adoption of Large Language Models (LLMs), concerns about potential misuse have emerged. To this end, watermarking has been adapted to LLM, enabling a simple and effective way to detect and monitor generated text. However, while the existing methods can differentiate between watermarked and unwatermarked text with high accuracy, they often face a trade-off between the quality of the generated text and the effectiveness of the watermarking process. In this work, we present a novel type of LLM watermark, Sparse WatermARK (or SpARK), which aims to mitigate this trade-off by applying watermarks to a small subset of generated tokens distributed across the text. To demonstrate this type of watermark, we introduce two novel variants, SpARK-P and SpARK-R, which achieve sparsity by anchoring watermarked tokens to words that have specific Part-of-Speech (POS) tags and specific hash values w.r.t a pseudorandom hash function, respectively. Our experimental results demonstrate that the proposed watermarking schemes, albeit embarrassingly simple, are incredibly effective, achieving high detectability while generating text that outperforms previous LLM watermarking methods in quality across various tasks. SpARK further advances the watermarking capability for LLMs while maintaining their generated text quality.

Large language models (LLMs) have achieved impressive success across many NLP tasks, yet implicit discourse relation classification (IDRC) is still dominated by encoder-only pre-trained language models such as RoBERTa. This may be due to earlier reports that ChatGPT performs poorly on IDRC in zero-shot settings. In this paper, we show that fine-tuned LLMs can perform on par with, or even better than, existing encoder-based approaches. Nevertheless, we find that LLMs alone struggle to capture subtle lexical relations between arguments for the task. To address this, we propose a two-step strategy that enriches arguments with explicit lexical-level semantic cues before fine-tuning. Experiments demonstrate substantial gains, particularly in cross-domain scenarios, with F1 scores improved by more than 10 points compared to strong baselines.

Conversational recommender systems (CRS) have advanced with large language models, showing strong results in domains like movies. These domains typically involve fixed content and passive consumption, where user preferences can be matched by genre or theme. In contrast, games present distinct challenges: fast-evolving catalogs, interaction-driven preferences (e.g., skill level, mechanics, hardware), and increased risk of unsafe responses in open-ended conversation. We propose MATCHA, a multi-agent framework for CRS that assigns specialized agents for intent parsing, tool-augmented retrieval, multi-LLM ranking with reflection, explanation, and risk control which enabling finer personalization, long-tail coverage, and stronger safety. Evaluated on real user request dataset, MATCHA outperforms six baselines across eight metrics, improving Hit@5 by 20%, reducing popularity bias by 24%, and achieving 97.9% adversarial defense. Human and virtual-judge evaluations confirm improved explanation quality and user alignment. Code will be released upon acceptance.

Large Language Model (LLM)–based Multi-Agent Systems (MAS) have emerged as a powerful paradigm for tackling complex, multi-step tasks across diverse domains. However, despite their impressive capabilities, MAS remain susceptible to adversarial manipulation. Existing studies typically examine isolated attack surfaces or specific scenarios, leaving a lack of holistic understanding of MAS vulnerabilities. To bridge this gap, we introduce PEAR, a benchmark for systematically evaluating both the utility and vulnerability of planner–executor MAS. While compatible with various MAS architectures, our benchmark focuses on the planner–executor structure—a practical and widely adopted design. Through extensive experiments, we find that (1) a weak planner degrades overall clean task performance more severely than a weak executor; (2) while a memory module is essential for the planner, incorporating a memory module into the executor yields only marginal improvements in clean-task performance; (3) there exists a trade-off between task performance and robustness; and (4) attacks targeting the planner are particularly effective at misleading the system. These findings offer actionable insights for enhancing the robustness of MAS and lay the groundwork for principled defenses in multi-agent settings.

Schema linking—the process of aligning natural language questions with database schema elements—is a critical yet underexplored component of Text-to-SQL systems. While recent methods have focused primarily on improving SQL generation, they often neglect the retrieval of relevant schema elements, which can lead to hallucinations and execution failures. In this work, we propose a context-aware bidirectional schema retrieval framework that treats schema linking as a standalone problem. Our approach combines two complementary strategies: table-first retrieval followed by column selection, and column-first retrieval followed by table selection. It is further augmented with techniques such as question decomposition, keyword extraction, and keyphrase extraction. Through comprehensive evaluations on challenging benchmarks such as BIRD and Spider, we demonstrate that our method significantly improves schema recall while reducing false positives. Moreover, SQL generation using our retrieved schema consistently outperforms full-schema baselines and closely approaches oracle performance, all without requiring query refinement. Notably, our method narrows the performance gap between full and perfect schema settings by 50%. Our findings highlight schema linking as a powerful lever for enhancing Text-to-SQL accuracy and efficiency.

Human communication is often implicit, conveying tone, identity, and intent beyond literal meanings. While large language models have achieved strong performance on explicit tasks such as summarization and reasoning, their capacity for expressivity, or implicit communication, remains underexplored. We introduce ExpressivityBench, a framework for evaluating the expressivity of LLMs using information-theoretic communication models. Our approach quantifies how well LLM-generated text communicates target properties without explicit mention, across nine tasks spanning emotion, identity, and tone. To enable scalable and reproducible evaluation, we employ LLM-based graders validated against human judgments. Our results reveal that while models are adept at expressing affective content, they struggle with sociolinguistic signals, lagging behind human baselines. This study provides a necessary step to evaluate human-like implicit communication, with implications for applications such as education, mental health support, and socially-aware dialogue systems. We provide code and data for our benchmark alongside our paper.