EACL 2026 论文列表

Meta-reviewing is a pivotal stage in the peer-review process, serving as the final step in determining whether a paper is recommended for acceptance. Prior research on meta-reviewing has treated this as a summarization problem over review reports. However, complementary to this perspective, meta-reviewing is a decision-making process that requires weighing reviewer arguments and placing them within a broader context. Prior research has demonstrated that decision-makers can be effectively assisted in such scenarios via dialogue agents. In line with this framing, we explore the practical challenges for realizing dialog agents that can effectively assist meta-reviewers. Concretely, we first address the issue of data scarcity for training dialogue agents by generating synthetic data using Large Language Models (LLMs) based on a self-refinement strategy to improve the relevance of these dialogues to expert domains. Our experiments demonstrate that this method produces higher-quality synthetic data and can serve as a valuable resource towards training meta-reviewing assistants. Subsequently, we utilize this data to train dialogue agents tailored for meta-reviewing and find that these agents outperform off-the-shelf LLM-based assistants for this task. Finally, we apply our agents in real-world meta-reviewing scenarios and confirm their effectiveness in enhancing the efficiency of meta-reviewing.

Understanding fine-grained links between documents is crucial for many applications, yet progress is limited by the lack of efficient methods for for data curation. To address this limitation, we introduce a domain-agnostic framework for bootstrapping sentence-level cross-document links from scratch. Our approach (1) generates and validates semi-synthetic datasets of linked documents, (2) uses these datasets to benchmark and shortlist the best-performing linking approaches, and (3) applies the shortlisted methods in large-scale human-in-the-loop annotation of natural text pairs. We apply the framework in two distinct domains – peer review and news – and show that combining retrieval models with LLMs achieves a 73% human approval rate for suggested links, more than doubling the acceptance of strong retrievers alone. Our framework allows users to produce novel datasets that enable systematic study of cross-document understanding, supporting downstream tasks such as media framing analysis and peer review assessment. All code, data, and annotation protocols are released to facilitate future research.

Multiple-Choice Questions (MCQs) are often used to assess knowledge, reasoning abilities, and even values encoded in large language models (LLMs). While the effect of multilingualism has been studied on LLM factual recall, this paper seeks to investigate the less explored question of language-induced variation in value-laden MCQ responses. Are multilingual LLMs consistent in their responses across languages, i.e. behave like theoretical polyglots, or do they answer value-laden MCQs depending on the language of the question, like a multitude of monolingual models expressing different values through a single model? We release a new corpus, the Multilingual European Value Survey (MEVS), which, unlike prior work relying on machine translation or ad hoc prompts, solely comprises human-translated survey questions aligned in 8 European languages. We administer a subset of those questions to over thirty multilingual LLMs of various sizes, manufacturers and alignment-fine-tuning status under comprehensive, controlled prompt variations including answer order, symbol type, and tail character. Our results show that while larger, instruction-tuned models display higher overall consistency, the robustness of their responses varies greatly across questions, with certain MCQs eliciting total agreement within and across models while others leave LLM answers split. Language-specific behavior seems to arise in all consistent, instruction-fine-tuned models, but only on certain questions, warranting a further study of the selective effect of preference fine-tuning.

Despite their impressive performance, vision-language models (VLMs) still struggle on culturally situated inputs. To understand how VLMs process culturally grounded information, we study the presence of culture-sensitive neurons, i.e., neurons whose activations show preferential sensitivity to inputs associated with particular cultural contexts. We examine whether such neurons are important for culturally diverse visual question answering and where they are located. Using the CVQA benchmark, we identify neurons of culture selectivity and perform diagnostic tests by deactivating the neurons flagged by various identification methods. Experiments on three VLMs across 25 cultural groups demonstrate the existence of neurons whose ablation disproportionately harms performance on questions about the corresponding cultures, while having limited effects on others. Moreover, we introduce a new margin-based selector—Contrastive Activation Margin (ConAct)—and show that it outperforms probability- and entropy-based methods in identifying neurons associated with cultural selectivity. Finally, our layer-wise analyses reveal that such neurons are not uniformly distributed: they cluster in specific decoder layers in a model-dependent way.

Large language models (LLMs) are increasingly adopted for handling structured data, including tabular and relational inputs, despite mostly being pretrained on unstructured text. This raises a key question: how effectively do pretrained representations from language-focused LLMs transfer to tasks involving structured inputs? We address this through controlled experiments using two small open-source LLMs, systematically re-initializing subsets of layers with random weights before fine-tuning on structured datasets and comparing results to unstructured datasets. Our analyses show that, for structured data, most pretrained depth contributes little, with performance often saturating after the first few layers, whereas unstructured tasks benefit more consistently from deeper pretrained representations. Pretraining remains useful mainly in low-resource settings, with its impact diminishing as more training data becomes available.

Large Language Models (LLMs) are integral to applications such as conversational agents and content creation, where precise control over a model’s personality is essential for maintaining tone, consistency, and user engagement. However, prevailing prompt-based or fine-tuning approaches either lack robustness or demand large-scale training data, making them costly and impractical. In this paper, we present PALETTE (Personality Adjustment by LLM SElf-TargeTed quEries), a novel method for personality editing in LLMs. Our approach introduces adjustment queries, where self-referential statements grounded in psychological constructs are treated analogously to factual knowledge, enabling direct editing of personality-related responses. Unlike fine-tuning, PALETTE requires only 12 editing samples to achieve substantial improvements in personality alignment across personality dimensions. Experimental results from both automatic and human evaluations demonstrate that our method enables more stable and well-balanced personality control in LLMs.

We present BabyBabelLM, a multilingual collection of datasets modeling the language a person observes from birth until they acquire a native language. We curate developmentally plausible pretraining data aiming to cover the equivalent of 100M English words of content in each of 45 languages. We compile evaluation suites and train baseline models in each language. BabyBabelLM aims to facilitate multilingual pretraining and cognitive modeling.

The growing collaboration between humans and AI models in generative tasks has introduced new challenges in distinguishing between human-written, LLM-generated, and human-LLM collaborative texts. In this work, we collect a multilingual, multi-domain, multi-generator dataset FAIDSet. We further introduce a fine-grained detection framework FAID to classify text into these three categories, and also to identify the underlying LLM family of the generator. Unlike existing binary classifiers, FAID is built to capture both authorship and model-specific characteristics. Our method combines multi-level contrastive learning with multi-task auxiliary classification to learn subtle stylistic cues. By modeling LLM families as distinct stylistic entities, we incorporate an adaptation to address distributional shifts without retraining for unseen data. Our experimental results demonstrate that FAID outperforms several baselines, particularly enhancing the generalization accuracy on unseen domains and new LLMs, thus offering a potential solution for improving transparency and accountability in AI-assisted writing. Our data and code are available at https://github.com/mbzuai-nlp/FAID.

Recent studies suggest that transformer-based vision-language models (VLMs) capture the multimodality of concept processing in the human brain. However, a systematic evaluation exploring different types of VLM architectures and the role played by visual and textual context is still lacking. Here, we analyse multiple VLMs employing different strategies to integrate visual and textual modalities, along with language-only counterparts. We measure the alignment between concept representations by models and existing (fMRI) brain responses to concept words presented in two experimental conditions, where either visual (pictures) or textual (sentences) context is provided. Our results reveal that VLMs outperform the language-only counterparts in both experimental conditions. However, controlled ablation studies show that only for some VLMs, such as LXMERT and IDEFICS2, brain alignment stems from genuinely learning more human-like concepts during pretraining, while others are highly sensitive to the context provided at inference. Additionally, we find that vision-language encoders are more brain-aligned than more recent, generative VLMs. Altogether, our study shows that VLMs align with human neural representations in concept processing, while highlighting differences among architectures. We open-source code and materials to reproduce our experiments at: https://github.com/dmg-illc/vl-concept-processing.

Sparse Autoencoders (SAEs) are powerful tools for interpreting neural representations, yet their use in audio remains underexplored. We train SAEs across all encoder layers of Whisper and HuBERT, provide an extensive evaluation of their stability, interpretability, and show their practical utility. Over 50% of the features remain consistent across random seeds, and reconstruction quality is preserved. SAE features capture general acoustic and semantic information as well as specific events, including environmental noises and paralinguistic sounds (e.g. laughter, whispering) and disentangle them effectively, requiring removal of only 19-27% of features to erase a concept. Feature steering reduces Whisper’s false speech detections by 70% with negligible WER increase, demonstrating real-world applicability. Finally, we find SAE features correlated with human EEG activity during speech perception, indicating alignment with human neural processing. The code and checkpoints are available at https://github.com/audiosae/audiosae_demo.

In the global drive toward carbon neutrality, deeply coordinated smart energy systems underpin industrial transformation, yet their interdisciplinary, fragmented, and fast-evolving expertise prevents general-purpose LLMs, lacking domain knowledge and physical-constraint awareness, from delivering precise engineering-aligned inference and generation. To address these challenges, we introduce Helios, the first large language model tailored to the smart energy domain, together with a comprehensive suite of resources to advance LLM research in this field. Specifically, we develop Enersys, a multi-agent collaborative framework for end-to-end dataset construction, through which we produce: (1) the first smart energy knowledge base, EnerBase, to enrich the model’s foundational expertise; (2) the first instruction fine-tuning dataset, EnerInsruct, to strengthen performance on domain-specific downstream tasks; and (3) the first RLHF dataset, EnerReinforce, to align the model with human preferences and industry standards. Leveraging these resources, Helios undergoes large-scale pretraining, SFT, and RLHF. We also release EnerBench, the first benchmark for evaluating LLMs in smart energy scenarios, and demonstrate that our approach significantly enhances domain knowledge mastery, task execution accuracy, and alignment with human preferences.

Retrieval-augmented generation (RAG) is a prevalent approach for building LLM-based question-answering systems that can take advantage of external knowledge databases. Due to the complexity of real-world RAG systems, there are many potential causes for erroneous outputs. Understanding the range of errors that can occur in practice is crucial for robust deployment. We present a new taxonomy of the error types that can occur in realistic RAG systems, examples of each, and practical advice for addressing them. Additionally, we curate a dataset of erroneous RAG responses annotated by error types. We then propose an auto-evaluation method aligned with our taxonomy that can be used in practice to track and address errors during development. Code and data are available at https://github.com/layer6ai-labs/rag-error-classification.

Does the prior knowledge of the vision encoder constrain the capability boundary of Multi-modal Large Language Models (MLLMs)? While most existing research treats MLLMs as unified systems optimized through end-to-end training, the impact of vision encoder’s prior knowledge is seldom investigated. In this work, we introduce a novel metric Rank_e to quantify the effect of prior knowledge of the vision encoder on MLLM performance. Our analysis reveals a positive correlation between prior knowledge and MLLM performance. Moreover, we find that domain-specific fine-tuning using solely end-to-end visual question answering (VQA) data is insufficient, particularly for entities with low inherent visual prior knowledge. To address this issue, we propose VisPRE (Vision Prior Remediation), a two-stage training framework that explicitly incorporates prior knowledge at the vision encoder level. Experimental results demonstrate that augmenting vision encoder’s prior knowledge substantially boosts the visual understanding capabilities of MLLMs, offering a novel and effective strategy for improving performance, especially in scenarios involving uncommon visual entities.

Training Large Language Models (LLMs) with high multilingual coverage is becoming increasingly important — especially when monolingual resources are scarce. Recent studies have found that LLMs process multilingual inputs in shared concept spaces, thought to support generalization and cross-lingual transfer. However, these prior studies often do not use causal methods, lack deeper error analysis or focus on the final model only, leaving open how these spaces emerge during training. We investigate the development of language-agnostic concept spaces during pretraining of EuroLLM through the causal interpretability method of activation patching. We isolate cross-lingual concept representations, then inject them into a translation prompt to investigate how consistently translations can be altered, independently of the language. We find that shared concept spaces emerge early and continue to refine, but that alignment with them is language-dependent. Furthermore, in contrast to prior work, our fine-grained manual analysis reveals that some apparent gains in translation quality reflect shifts in behavior — like selecting senses for polysemous words or translating instead of copying cross-lingual homographs — rather than improved translation ability. Our findings offer new insight into the training dynamics of cross-lingual alignment and the conditions under which causal interpretability methods offer meaningful insights in multilingual contexts.

The intentional creation and spread of disinformation poses a significant threat to public discourse. However, existing English datasets and research rarely address the intentionality behind the disinformation. This work presents MALINT, the first human-annotated English corpus developed in collaboration with expert fact-checkers to capture disinformation and its malicious intent. We utilize our novel corpus to benchmark 12 language models, including small language models (SLMs) such as BERT and large language models (LLMs) like Llama 3.3, on binary and multilabel intent classification tasks. Moreover, inspired by inoculation theory from psychology and communication studies, we investigate whether incorporating knowledge of malicious intent can improve disinformation detection. To this end, we propose intent-based inoculation, an intent-augmented reasoning for LLMs that integrates intent analysis to mitigate the persuasive impact of disinformation. Analysis on six disinformation datasets, five LLMs, and seven languages shows that intent-augmented reasoning improves zero-shot disinformation detection. To support research in intent-aware disinformation detection, we release the MALINT dataset with annotations from each annotation step.

Large language models (LLMs) increasingly rely on external tools and APIs to execute complex tasks specified in natural language. Evaluating such tool-calling capabilities in realistic enterprise settings is challenging: APIs are often proprietary, heterogeneous, and difficult to share, limiting reproducible benchmarks. To address this, we introduce Live API Bench, a comprehensive benchmark constructed by transforming NL2SQL datasets into interactive API environments. Our pipeline converts SQL queries from BIRD-SQL into executable API sequences across three formulations—SLOT, SEL, and REST—covering minimal general-purpose operations, domain-specific multi-step tasks, and function-oriented RESTful interactions, respectively. The benchmark spans 11 databases with over 2,500 invocable tools, paired with human-authored queries, ground-truth API sequences, and verified final answers. Live API Bench enables systematic evaluation of core challenges in tool use, including error handling, sequential reasoning, parameter generation, response parsing, and robustness across diverse domains. We evaluate 10 LLMs and 4 ReACT agents, observing low task completion rates (7–47%), which improve modestly to 50% under interactive agent settings, highlighting substantial scope for improving LLM tool-calling performance. We release all code and data associated with this paper.

Automated Essay Scoring (AES) has gained increasing attention in recent years, yet research on Arabic AES remains limited due to the lack of publicly available datasets. To address this, we introduce LAILA, the largest publicly available Arabic AES dataset to date, comprising 7,859 essays annotated with holistic and trait-specific scores on seven dimensions: relevance, organization, vocabulary, style, development, mechanics, and grammar. We detail the dataset design, collection, and annotations, and provide benchmark results using state-of-the-art Arabic and English models in prompt-specific and cross-prompt settings. LAILA fills a critical need in Arabic AES research, supporting the development of robust scoring systems.

Accurately forecasting large stock price movements after corporate earnings announcements is a longstanding challenge. Existing methods–sentiment lexicons, fine-tuned encoders, and standalone LLMs–often lack temporal-causal reasoning and are prone to narrative bias, echoing overly optimistic managerial tone. We introduce Context-Enriched Agentic RAG (CARAG), a retrieval-augmented framework that deploys a team of cooperative LLM agents, each specializing in a distinct analytical task: evaluating historical performance, assessing the credibility of guidance, or benchmarking against peers.Agents retrieve structured evidence from a Causal-Temporal Knowledge Graph (CTKG) built from financial statements and earnings calls, enabling grounded, context-rich reasoning. This design mitigates LLM hallucinations and produces more objective predictions.Without task-specific training, our system achieves state-of-the-art zero-shot performance across NASDAQ, NYSE, and MAEC datasets, outperforming both larger LLMs and fine-tuned models in macro-F1, MCC, and Sharpe, beating market benchmarks (S P 500 and Nasdaq) for the same forecasting horizon. Code, datasets, prompts, and implementation details are included in the supplementary material to ensure full reproducibility.

Despite the state-of-the-art performance of Large Language Models (LLMs) achieved on many tasks, their massive scale often leads to high computational and environmental costs, limiting their accessibility. Parameter-Efficient Fine-Tuning (PEFT) methods address this challenge by reducing the number of trainable parameters while maintaining strong downstream performance. Despite the advances in PEFT methods, current evaluations remain limited (in terms of evaluated models and datasets) and difficult to reproduce. To bridge this gap, we introduce PEFT-Bench, a unified end-to-end benchmark for evaluating diverse PEFT methods on autoregressive LLMs. We demonstrate its usage across 27 NLP datasets and 7 PEFT methods. To account for different PEFT training and inference factors, we also introduce the PEFT Soft Cost Penalties (PSCP) metric, which takes trainable parameters, inference speed, and training memory usage into account.