EACL 2026 论文列表

Prompt optimization (PO) provides a practical way to improve response quality when users lack the time or expertise to manually craft effective prompts. Existing methods typically rely on LLMs’ self-generation ability to optimize prompts. However, due to limited downward compatibility, the instruction-heavy prompts generated by advanced LLMs can overwhelm lightweight inference models and degrade response quality, while also lacking interpretability due to implicit optimization. In this work, we rethink prompt optimization through the lens of explicit and interpretable design. We first identify a set of model-agnostic prompt quality merits and empirically validate their effectiveness in enhancing prompt and response quality. We then introduce MePO, a merit-guided, locally deployable prompt optimizer trained on our merit-guided prompt preference dataset generated by a lightweight LLM. MePO avoids online optimization, reduces privacy concerns, and, by learning clear, interpretable merits, generalizes effectively to both large-scale and lightweight inference models. Experiments demonstrate that MePO achieves better results across diverse tasks and model types, offering a scalable and robust solution for real-world deployment. The code, model and dataset can be found in https://github.com/MidiyaZhu/MePO.

Continual learning in natural language processing plays a crucial role in adapting to evolving data and preventing catastrophic forgetting. Despite significant progress, existing methods still face challenges, such as inefficient parameter reuse across tasks, risking catastrophic forgetting when tasks are dissimilar, and the unnecessary introduction of new parameters for each task, which hampers knowledge sharing among similar tasks. To tackle these issues, we propose a Sparse Adapter Fusion Method (SAFM), which dynamically fuses old and new adapters to address these challenges. SAFM operates in two stages: the decision stage and the tuning stage. In the decision stage, SAFM determines whether to incorporate a new adapter, reuse an existing one, or add an empty adapter. The architecture search procedure, designed to prioritize reusing or adding empty adapters, minimizes parameter consumption and maximizes reuse. In the tuning stage, SAFM especially facilitates a layer-wise loss to encourage differentiation between adapters, effectively capturing knowledge within the same task. Experimental results consistently show that SAFM outperforms state-of-the-art (SOTA) methods, achieving comparable performance while utilizing less than 60% of the parameters.

We consider coverless steganography where a Large Language Model (LLM) is used to generate stego-texts in combination with arithmeticic coding. An efficient method should embed secret bits in as few language tokens as possible while keeping the stego-text as natural as possible. We show that this problem is equivalent to maximizing the entropy of a replacement probability distribution of the next token generation, subject to a constraint on the divergence between the new distribution and the original one produced by the LLM. A closed-form solution is provided under either the KL divergence or the total variation constraint. Several important practical issues are also tackled: 1) An often-overlooked tokenization mismatch issue is resolved with a simple prompt selection approach, 2) The combination of the optimized distribution and the vocabulary truncation technique is considered, and 3) The incorporation of the proposed approach with existing (potentially non arithemtic coding based) techniques, e.g., the Discop technique.

We expose a critical limitation in current approaches to machine unlearning in language models: despite the apparent success of unlearning algorithms, information about the forgotten data remains linearly decodable from internal representations. To systematically assess this discrepancy, we introduce an interpretable, information-theoretic framework for auditing unlearning using Partial Information Decomposition (PID). By comparing model representations before and after unlearning, we decompose the mutual information with the forgotten data into distinct components, formalizing the notions of unlearned and residual knowledge. Our analysis reveals that redundant information, shared across both models, constitutes residual knowledge that persists post-unlearning and correlates with susceptibility to known adversarial reconstruction attacks. Leveraging these insights, we propose a representation-based risk score that can guide abstention on sensitive inputs at inference time, providing a practical mechanism to mitigate privacy leakage. Our work introduces a principled, representation-level audit for unlearning, offering theoretical insight and actionable tools for safer deployment of language models.

While large language models (LLMs) are generally considered proficient in generating language, how similar their language usage is to that of humans remains understudied. In this paper, we test whether models exhibit linguistic convergence, a core pragmatic element of human language communication: do models adapt, or converge, to the linguistic patterns of their user? To answer this, we systematically compare model completions of existing dialogues to original human responses across sixteen language models, three dialogue corpora, and various stylometric features. We find that models strongly converge to the conversation’s style, often significantly overfitting relative to the human baseline. While convergence patterns are often feature-specific, we observe consistent shifts in convergence across modeling settings, with instruction-tuned and larger models converging less than their pretrained and smaller counterparts. Given the differences in human and model convergence patterns, we hypothesize that the underlying mechanisms driving these behaviors are very different.

Social media platforms face escalating challenges in detecting harmful content that promotes muscle dysmorphic behaviors and cognitions (bigorexia). This content can evade moderation by camouflaging as legitimate fitness advice and disproportionately affects adolescent males. We address this challenge with BigTokDetect, a clinically informed framework for identifying pro-bigorexia content on TikTok. We introduce BigTok, the first expert-annotated multimodal benchmark dataset of over 2,200 TikTok videos labeled by clinical psychiatrists across five categories and eighteen fine-grained subcategories. Comprehensive evaluation of state-of-the-art vision-language models reveals that while commercial zero-shot models achieve the highest accuracy on broad primary categories, supervised fine-tuning enables smaller open-source models to perform better on fine-grained subcategory detection. Ablation studies show that multimodal fusion improves performance by 5 to 15 percent, with video features providing the most discriminative signals. These findings support a grounded moderation approach that automates detection of explicit harms while flagging ambiguous content for human review, and they establish a scalable framework for harm mitigation in emerging mental health domains.

Categorization is a core component of human linguistic competence. We investigate how a transformer-based language model (LM) learns linguistic categories by comparing its behaviour over the course of training to behaviours which characterize abstract feature–based and concrete exemplar–based accounts of human language acquisition. We investigate how lexical semantic and syntactic categories emerge using novel divergence-based metrics that track learning trajectories using next-token distributions. In experiments with GPT-2 small, we find that (i) when a construction is learned, abstract class-level behaviour is evident at earlier steps than lexical item–specific behaviour, and (ii) that different linguistic behaviours emerge abruptly in sequence at different points in training, revealing that abstraction plays a key role in how LMs learn. This result informs the models of human language acquisition that LMs may serve as an existence proof for.

Attention regularisation aims to supervise the attention patterns in language models like BERT. Various studies have shown that using human-annotated rationales, in the form of highlights that explain why a text has a specific label, can have positive effects on model generalisability. In this work, we ask to what extent attention regularisation with human-annotated rationales improve model performance and model robustness, as well as susceptibility to spurious correlations. We compare regularisation on human rationales with randomly selected tokens, a baseline which has hitherto remained unexplored.Our results suggest that often, attention regularisation with randomly selected tokens yields similar improvements to attention regularisation with human-annotated rationales. Nevertheless, we find that human-annotated rationales surpass randomly selected tokens when it comes to reducing model sensitivity to strong spurious correlations.

Recent advances in reinforcement learning with verifiable rewards (RLVR) show that large language models enhance their reasoning abilities when trained with verifiable signals. However, due to reward sparsity, effectiveness depends heavily on selecting samples of appropriate difficulty. In this work, we present a formal analysis of online difficulty-aware filtering and establish its theoretical foundations. We show that expected policy improvement is lower-bounded by the variance of task-level success probabilities, implying that selecting tasks of intermediate difficulty maximizes learning efficiency. Building on this, we demonstrate that balanced filtering maximizes this lower bound, leading to superior performance and sample efficiency. Evaluations across multiple math reasoning benchmarks validate that balanced filtering consistently enhances convergence speed and final performance, achieving up to +12% gains in less than half the training steps of standard GRPO. By extending our analysis to various reward distributions, we provide a principled foundation for future RLVR curriculum strategies, confirmed through both theoretical analysis and extensive empirical results.

Large language models (LLMs) often respond confidently to questions even when they lack the necessary information, leading to hallucinated answers. In this work, we study the problem of (un)answerability detection, focusing on extractive question answering (QA) where the model should determine if a passage contains sufficient information to answer a given question. We propose a simple approach for identifying a direction in the model’s activation space that captures unanswerability and uses it for classification. This direction is selected by applying activation additions during inference and measuring their impact on the model’s abstention behavior. We show that projecting hidden activations onto this direction yields a reliable score for (un)answerability classification. Experiments on two open-weight LLMs and four extractive QA benchmarks show that our method effectively detects unanswerable questions and generalizes better across datasets than existing prompt-based and classifier-based approaches. Moreover, the obtained directions extend beyond extractive QA to unanswerability that stems from factors, such as lack of scientific consensus and subjectivity. Last, causal interventions show that adding or ablating the directions effectively controls the abstention behavior of the model.

Automated feature engineering (AutoFE) liberates data scientists from the burden of manual feature construction. The semantic information of datasets contains rich context information for feature engineering but has been underutilized in many existing AutoFE works. We present PromptFE, a novel AutoFE framework that leverages large language models (LLMs) to automatically construct features in a compact string format and generate semantic explanations based on dataset descriptions. By learning the performance of constructed features in context, the LLM iteratively improves feature construction. We demonstrate through experiments on real-world datasets the superior performance of PromptFE over state-of-the-art AutoFE methods. We verify the impact of dataset semantic information and provide comprehensive study on the LLM-based feature construction process.

As Large Language Models (LLMs) become increasingly integrated into everyday life and information ecosystems, concerns about their implicit biases continue to persist. While prior work has primarily examined socio-demographic and left–right political dimensions, little attention has been paid to how LLMs align with broader geopolitical value systems, particularly the democracy–authoritarianism spectrum. In this paper, we propose a novel methodology to assess such alignment, combining (1) the F-scale, a psychometric tool for measuring authoritarian tendencies, (2) FavScore, a newly introduced metric for evaluating model favorability toward world leaders, and (3) role-model probing to assess which figures are cited as general role models by LLMs. We find that LLMs generally favor democratic values and leaders, but exhibit increased favorability toward authoritarian figures when prompted in Mandarin. Further, models are found to often cite authoritarian figures as role models, even outside explicitly political contexts. These results shed light on ways LLMs may reflect and potentially reinforce global political ideologies, highlighting the importance of evaluating bias beyond conventional socio-political axes.

Open-world Question Answering (OW-QA) over knowledge graphs (KGs) aims to answer questions over incomplete or evolving KGs. Traditional KGQA assumes a closed world where answers must exist in the KG, limiting real-world applicability. In contrast, open- world QA requires inferring missing knowledge based on graph structure and context. Large language models (LLMs) excel at language understanding but lack structured reasoning. Graph neural networks (GNNs) model graph topology but struggle with semantic interpretation. Existing systems integrate LLMs with GNNs or graph retrievers. Some support open-world QA but rely on structural embeddings without semantic grounding. Most assume observed paths or complete graphs, making them unreliable under missing links or multi-hop reasoning. We present GLOW, a hybrid system that combines a pre-trained GNN and an LLM for open-world KGQA. The GNN predicts top-k candidate answers from the graph structure. These, along with relevant KG facts, are serialized into a structured prompt (e.g., triples and candidates) to guide the LLM’s reasoning. This enables joint reasoning over symbolic and semantic signals, without relying on retrieval or fine-tuning. To evaluate generalization, we introduce GLOW-BENCH, a 1,000-question benchmark over incomplete KGs across diverse domains. GLOW outperforms existing LLM–GNN systems on standard benchmarks and GLOW-BENCH, achieving up to 53.3% and an average 38% improvement.

Large language models (LLMs) show promise in offering emotional support and generating empathetic responses for individuals in distress, but their ability to deliver culturally sensitive support remains underexplored due to a lack of resources. In this work, we introduce , the first dataset designed for this task, spanning four cultures and including 1,729 distress messages, 1,523 cultural signals, and 1,041 support strategies with fine-grained emotional and cultural annotations. Leveraging , we (i) develop and test four adaptation strategies for guiding three state-of-the-art LLMs toward culturally sensitive responses; (ii) conduct comprehensive evaluations using LLM-as-a-Judge, in-culture human annotators, and clinical psychologists; (iii) show that adapted LLMs outperform anonymous online peer responses, and that simple cultural role-play is insufficient for cultural sensitivity; and (iv) explore the application of LLMs in clinical training, where experts highlight their potential in fostering cultural competence in novice therapists.

Text Generation has achieved remarkable performance using large language models. It has also been recently well-studied that these large language models are capable of creative generation tasks but prominently for high-resource languages. This prompts a fundamental question: \textit{Is there a way to utilize these (large) language models for structured poetry generation in a low-resource language, such as Sanskrit?} We present Chandomitra, an English input to structured Sanskrit Poetry translation dataset, specifically adhering to the Anushtubh meter. We benchmark various open and closed models, and scrutinize specialized techniques such as constrained decoding and instruction fine-tuning, for the proposed task. Our constrained decoding methodology achieves 99.86% syntactic accuracy in generating metrically valid Sanskrit poetry, outperforming GPT-4o (1-shot: 31.24%). Our best-performing instruction-tuned model, on the other hand, performs better in semantic coherence with the English input, at the expense of slightly lower syntactic accuracy. Human evaluation further reveals that instruction fine-tuned model is better able to capture the poetic aspects.

In this paper, we present the first cross-lingual dataset that captures a transnational cultural phenomenon, focusing on the Chinese and Japanese “Jirai” subculture and its association with risky health behaviors. Our dataset of more than 15,000 annotated social media posts forms the core of JiraiBench, a benchmark designed to evaluate LLMs on culturally specific content. This unique resource allowed us to uncover an unexpected cross-cultural transfer in which Japanese prompts better handle Chinese content, indicating that cultural context can be more influential than linguistic similarity. Further evidence suggests potential cross-lingual knowledge transfer in fine-tuned models. This work proves the indispensable role of developing culturally informed, cross-lingual datasets for creating effective content moderation tools that can protect vulnerable communities across linguistic borders.

Large language models achieve strong performance on many tasks, but their training makes it hard to see which properties of the input support efficient linguistic rule learning.We ask how three cognitively-inspired principles of input design support sample-efficient linguistic rule induction: analogical structure, contrastive learning, and minimal contextual cue. We also ask how their effects compare to those of LLMs on the same controlled tasks.We implement these principles in structured sentence completion tasks that test English verb alternations.Lightweight models trained on hundreds to one-thousand such examples learn the alternation rules with high F1 on these tasks. Ablation studies show that analogical organisation is the main driver of sample efficiency, and contrastive distractors and minimal context help further gains. We also evaluate zero- and few-shot LLMs on the same tasks. In this controlled setting, the lightweight models reach higher F1 with far fewer task-specific data.We treat this contrast as a comparison between learning regimes rather than a general verdict on LLMs.Our results show that careful input organisation supports sample-efficient learning of linguistic rules and reveals distinct learning signatures for trained lightweight models and prompted LLMs.

Understanding user intentions is challenging for online platforms. Recent work on intention knowledge graphs addresses this but often lacks focus on connecting intentions, which is crucial for modeling user behavior and predicting future actions. This paper introduces a framework to automatically generate an intention knowledge graph, capturing connections between user intentions. Using the Amazon m2 dataset, we construct an intention graph with 351 million edges, demonstrating high plausibility and acceptance. Our model effectively predicts new session intentions and enhances product recommendations, outperforming previous state-of-the-art methods and showcasing the approach’s practical utility.

Antimicrobial resistance is a growing global health threat, driving interest in nanoparticle-based alternatives to conventional antibiotics. Inorganic nanoparticles (NPs) with intrinsic antibacterial properties show significant promise; however, efficiently identifying relevant studies from the rapidly expanding literature remains a major challenge. This step is crucial for enabling computational approaches that aim to model and predict NP efficacy based on physicochemical and structural features. In this study, we explore the effectiveness of traditional machine learning and deep learning methods in classifying scientific abstracts in the domain of NP-based antimicrobial research. We introduce the “Antibacterial Inorganic NAnoparticles Dataset” AINA of 7,910 articles, curated to distinguish intrinsic antibacterial NPs from studies focusing on drug carriers or surface-bound applications. Our comparative evaluation shows that a fine-tuned BioBERT classifier achieved the highest macro F1 (0.82), while a lightweight SVM model with TF-IDF features remained competitive (0.78), highlighting their utility in low-resource settings. AINA enables reproducible, large-scale identification of intrinsically bactericidal inorganic NPs. By reducing noise from non-intrinsic contexts, this work provides a foundation for mechanism-aware screening, database construction, and predictive modeling in antimicrobial NP research.