EACL 2026 论文列表

The advancement of reasoning capabilities in Large Language Models (LLMs) requires substantial amounts of high-quality reasoning data, particularly in mathematics. Existing data synthesis methods, such as data augmentation from annotated training sets or direct question generation based on relevant knowledge points and documents, have expanded datasets but face challenges in mastering the internal logic of the problem during generation and ensuring the verifiability of the solutions. To address these issues, we propose RV-Syn, a novel Rational and Verifiable mathematical Synthesis approach. RV-Syn first constructs a structured library of mathematical operations and then composes them into executable computational graphs, which serve as verifiable solution blueprints. These graphs are subsequently back-translated into complex problems, enabling solution-guided, logic-aware problem generation while inherently ensuring the verifiability of the solving process. Experimental results show RV-Syn surpasses existing synthesis methods, including those involving human-crafted problems. Our method achieves a 6.3% performance gain over the previous state-of-the-art synthetic data on LLaMA-3-8B and demonstrates superior data efficiency, outperforming others with only half the training data (50k vs. 100k), enabling a more scalable and robust reasoning dataset generation framework.

Recent advances in large language models (LLMs) have introduced explicit reasoning capabilities, yet the factors that truly drive their improved performance remain unclear. In this work, we disentangle the effects of reasoning quality and sequence length by fine-tuning 8B models on several Polish variants of the Mixture-of-Thoughts (MoT-PL) dataset, each representing a distinct reasoning style: Detailed, Summarized, BabyThink, Lengthy. We found that the model trained on high-quality reasoning traces achieved better average performance than all other models; neither longer reasoning with similar quality nor low-quality reasoning with similar length achieved similar gains. Qualitative and quantitative analyses further reveal that reasoning clarity, rather than verbosity, is the dominant factor driving model performance. These findings underscore the importance of reasoning content quality in LLM training and provide new insights into designing more effective reasoning-oriented datasets and models.

Structured output from large language models (LLMs) has enhanced efficiency in processing generated information and is increasingly adopted in industrial applications. Prior studies have investigated the impact of structured output on LLMs’ generation quality, often presenting one-way findings. Some suggest that structured format enhances completeness and factual accuracy, while others argue that it restricts the reasoning capacity of LLMs and leads to reductions in standard evaluation metrics. Potential limitations of these assessments include restricted testing scenarios, weakly controlled comparative settings, and reliance on coarse metrics. In this work, we present a refined analysis using causal inference. Based on one assumed and two guaranteed constraints, we derive five potential causal structures characterizing the influence of structured output on LLMs’ generation: (1) collider without m-bias, (2) collider with m-bias, (3) single cause from instruction, (4) single cause from output format, and (5) independence. Across seven public and one developed reasoning tasks, we find that coarse metrics report positive, negative, or neutral effects of structured output on GPT-4o’s generation. However, causal inference reveals no causal impact in 43 out of 48 scenarios. In the remaining 5, 3 involve multifaceted causal structures influenced by concrete instructions. Further experiments show that OpenAI-o3 are more resilient to output formats than general-purpose GPT-4o and GPT-4.1, highlighting an unaware advantage of reasoning models.

Large Language Models (LLMs) have revolutionized inference across diverse natural language tasks, with larger models performing better but at higher computational costs. We propose a confidence-driven strategy that dynamically selects the most suitable model based on confidence estimates. By assessing a model’s confidence in handling the task and response accuracy, tasks that are likely to be solved correctly are retained, while more uncertain or complex cases are delegated to a larger model, ensuring reliability while minimizing computation. Specifically, we evaluate a model’s likelihood of knowing the correct answer and the probability that its response is accurate.Experiments on the Massive Multitask Language Understanding (MMLU) benchmark show that our approach achieves accuracy comparable to the largest model while reducing computational costs by 20% to 40%. When applied to GPT-4o API calls, it reduces token usage by approximately 60%, further improving cost efficiency. These findings indicate the potential of confidence-based model selection to enhance real-world LLM deployment, particularly in resource-constrained settings such as edge devices and commercial API applications.

We conduct a systematic audit of three widely used social reasoning benchmarks, SocialIQa, FauxPas-EAI, and ToMi, and uncover pervasive flaws in both benchmark items and evaluation methodology. Using five LLMs (GPT-3, 3.5, 4, o1, and LLaMA 3.1) as diagnostic tools, we identify structural, semantic, and pragmatic issues in benchmark design (e.g., duplicated items, ambiguous wording, and implausible answers), as well as scoring procedures that prioritize output form over the reasoning process. Through systematic human annotation and re-evaluation on cleaned benchmark subsets, we find that model scores often improve not due to due to erratic surface wording variations and not to improved reasoning. In fact, further analyses show that model performance is highly sensitive to minor input variations such as context availability and phrasing, revealing that high scores may reflect alignment with format-specific cues rather than consistent inference based on the input. These findings challenge the validity of current benchmark-based claims about social reasoning in LLMs, and highlight the need for evaluation protocols that assess reasoning as a process of drawing inference from available information, rather than as static output selection. We release audited data and evaluation tools to support more interpretable and diagnostic assessments of model reasoning

Federated fine-tuning of foundation models is impeded by the need to communicate billions of parameters. Low-rank adaptation (LoRA) alleviates this by updating only compact adapter matrices. However, varying client device capabilities lead to different adapter ranks, causing rank heterogeneity that undermines aggregation, and existing reconciliation methods still incur bias or inefficiency. To address this challenge, we propose RB-LoRA, a principled rank-balanced aggregation framework that decomposes each update into rank-wise components and aligns them using analytically derived weights. Experiments on both language and vision models demonstrate consistent improvements under one and three rounds of communication in federated learning.

Recent advances in Multimodal Large Language Models (MLLMs) have improved image recognition and reasoning, but video-related tasks remain challenging due to memory constraints from dense frame processing. Existing Video Moment Retrieval (VMR) methodologies rely on sparse frame sampling, risking potential information loss, especially in lengthy videos. We propose SMORE (See MORE, store less), a framework that enhances memory efficiency while maintaining high information resolution. SMORE (1) uses query-guided captions to encode semantics aligned with user intent, (2) applies query-aware importance modulation to highlight relevant segments, and (3) adaptively compresses frames to preserve key content while reducing redundancy. This enables efficient video understanding without exceeding memory budgets. Experimental validation reveals that SMORE achieves state-of-the-art performance on QVHighlights, Charades-STA, and ActivityNet-Captions benchmarks.

Vietnam ranks among the top countries in terms of both internet traffic and online toxicity. As a result, implementing embedding models for recommendation and content control duties in applications is crucial. However, a lack of large-scale test datasets, both in volume and task diversity, makes it tricky for scientists to effectively evaluate AI models before deploying them in real-world, large-scale projects. To solve this important problem, we introduce a Vietnamese benchmark, VN-MTEB for embedding models, which we created by translating a large number of English samples from the Massive Text Embedding Benchmark using our new automated framework, thereby contributing an extension of the Massive Multilingual Text Embedding Benchmark with our additional Vietnamese tasks and datasets. We leverage the strengths of large language models (LLMs) and cutting-edge embedding models to conduct translation and filtering processes to retain high-quality samples, guaranteeing a natural flow of language and semantic fidelity while preserving named entity recognition (NER) and code snippets. Our comprehensive benchmark consists of 41 datasets from six tasks specifically designed for Vietnamese text embeddings. In our analysis, we find that bigger and more complex models using Rotary Positional Embedding outperform those using Absolute Positional Embedding in embedding tasks.

Recent advancements in LLMs have significantly improved mathematical problem-solving, with models like GPT-4 achieving human-level performance. However, proficiently solving mathematical problems differs fundamentally from effectively teaching mathematics. To bridge this gap, we introduce the Bi-GSM8K benchmark, a bilingual English-Korean dataset enriched with teacher solutions, student solutions, and annotations marking students’ initial errors. This dataset is designed to evaluate two core capabilities of LLMs: (1) measuring similarity between student and teacher solutions, and (2) identifying the initial error point in student solutions. Our method achieves high agreement with human judgments, with Pearson 0.89 and Spearman 0.88 on English, and Pearson 0.89 and Spearman 0.87 on Korean. It also offers significantly lower latency and resource usage than commercial APIs, demonstrating strong computational efficiency. In the error detection task, open-source models achieved approximately 86% accuracy, with performance within 10% points of commercial LLMs API, suggesting strong practical potential. Our key contributions include the open-source release of Bi-GSM8K, novel evaluation metrics, and comparative analyses of LLM performance across languages.

Continual pre-training (CPT) has been widely adopted as a method for domain expansion in large language models. However, CPT has consistently been accompanied by challenges, such as the difficulty of acquiring large-scale domain-specific datasets and high computational costs. In this study, we propose a novel method called Test-Enhanced Learning for Language Model Enrichment (TELLME) to alleviate these issues. TELLME leverages the Test-Enhanced Learning (TEL) principle, whereby the model’s learning efficiency is improved using quizzes during training. It integrates this principle with CPT, thereby promoting efficient domain-specific knowledge acquisition and long-term memory retention. Experimental results demonstrate that TELLME outperforms existing methods by up to 23.6% in the financial domain and achieves a 9.8% improvement in long-term memory retention.

Information Retrieval (IR) methods aim to identify documents relevant to a query, which have been widely applied in various natural language tasks. However, existing approaches typically consider only the textual content within documents, overlooking the fact that documents can contain multiple modalities, including images and tables. Also, they often segment each long document into multiple discrete passages for embedding, which prevents them from capturing the overall document context and interactions between paragraphs. To address these two challenges, we propose a method that holistically embeds documents interleaved with multiple modalities by leveraging the capability of recent vision-language models that enable the processing and integration of text, images, and tables into a unified format and representation. Moreover, to mitigate the information loss from segmenting documents into passages, instead of representing and retrieving passages individually, we further merge the representations of segmented passages into one single document representation, while we additionally introduce a reranking strategy to decouple and identify the relevant passage within the document if necessary. Then, through extensive experiments on diverse IR scenarios considering both the textual and multimodal queries, we show that our approach substantially outperforms relevant baselines, thanks to the consideration of the multimodal information within documents.

With the increasing adoption of Large Language Models (LLMs) and Vision-Language Models (VLMs),rich document analysis technologies for applications like Retrieval-Augmented Generation (RAG)and visual RAG are gaining significant attention.Recent research indicates that using VLMs yields better RAG performance,but processing rich documents remains a challenge since a single page contains large amounts of information.In this paper, we present SCAN (SemantiC Document Layout ANalysis),a novel approach that enhances both textual and visual Retrieval-Augmented Generation (RAG) systemsthat work with visually rich documents.It is a VLM-friendly approach that identifies document components with appropriate semantic granularity,balancing context preservation with processing efficiency.SCAN uses a coarse-grained semantic approach that divides documents into coherent regions covering contiguous components.We trained the SCAN model by fine-tuning object detection models on an annotated dataset.Our experimental results across English and Japanese datasets demonstrate that applying SCAN improvesend-to-end textual RAG performance by up to 9.4 points and visual RAG performance by up to 10.4 points,outperforming conventional approaches and even commercial document processing solutions.

Large Language Models (LLMs) are increasingly being used to understand how scientific research evolves, drawing growing interest from the research community. However, limited work has explored the scientific fact-checking of research questions and claims from manuscripts, particularly within the NLP domain, an emerging direction for advancing scientific integrity and knowledge validation. In this work, we propose a novel scientific fact-checking dataset, SCINLP, tailored to the NLP domain. Our proposed framework on SCINLP systematically verifies the veracity of complex scientific research questions across varying rationale contexts, while also assessing their temporal positioning. SCINLP includes supporting and refuting research questions from a curated collection of influential and reputable NLP papers published between 2000 and 2024. In our framework, we use multiple LLMs and diverse rationale contexts from our dataset to examine scientific claims and research focus, complemented by feasibility judgments for deeper insight into scientific reasoning in NLP.

Recent multimodal large language models (MLLMs) extend language understanding beyond text to speech, enabling unified reasoning across modalities. While biases in text-based LLMs have been widely examined, their persistence and manifestation in spoken inputs remain underexplored. This work presents the first systematic investigation of speech bias in multilingual MLLMs.We construct and release the BiasInEar Dataset, a speech-augmented benchmark based on Global MMLU Lite, spanning English, Chinese, and Korean, balanced by gender and accent, and totaling 70.8 hours (\approx4,249 minutes) of speech with 11,200 questions. Using four complementary metrics (accuracy, entropy, APES, and Fleiss’ \kappa), we evaluate nine representative models under linguistic language and accent, demographic gender, and structural option order perturbations. Our findings reveal that MLLMs are relatively robust to demographic factors but highly sensitive to language and option order, suggesting that speech can amplify existing structural biases. Moreover, architectural design and reasoning strategy substantially affect robustness across languages. Overall, this study establishes a unified framework for assessing fairness and robustness in speech-integrated LLMs, bridging the gap between text- and speech-based evaluation.

Distilling advanced Large Language Models’ instruction-following capabilities into smaller models using a selected subset has become a mainstream approach in model training. While existing synthetic instruction data selection strategies rely mainly on single-dimensional signals (i.e., reward scores, model perplexity), they fail to capture the complexity of instruction-following across diverse fields. Therefore, we investigate more diverse signals to capture comprehensive instruction-response pair characteristics and propose three foundational metrics that leverage Multi-LLMs wisdom, informed by (1) diverse LLM responses and (2) reward model assessment. Building upon base metrics, we propose CrowdSelect, an integrated metric incorporating a clustering-based approach to maintain response diversity. Our comprehensive experiments demonstrate that our foundation metrics consistently improve performance across 4 base models on MT-bench and Arena-Hard. CrowdSelect, efficiently incorporating all metrics, achieves state-of-the-art performance in both Full and LoRA fine-tuning, showing improvements of 4.81% on Arena-Hard and 11.1% on MT-bench with Llama-3.2-3b-instruct. We hope our findings will bring valuable insights for future research in this direction.

Natural Language Processing (NLP) is integral to social media analytics but often processes content containing Personally Identifiable Information (PII), behavioral cues, and metadata raising privacy risks such as surveillance, profiling, and targeted advertising. To systematically assess these risks, we review 203 peer-reviewed papers and propose the NLP Privacy Risk Identification in Social Media (NLP-PRISM) framework, which evaluates vulnerabilities across six dimensions: data collection, preprocessing, visibility, fairness, computational risk, and regulatory compliance. Our analysis shows that transformer models achieve F1-scores ranging from 0.58–0.84, but incur a 1\% - 23\% drop under privacy-preserving fine-tuning. Using NLP-PRISM, we examine privacy coverage in six NLP tasks: sentiment analysis (16), emotion detection (14), offensive language identification (19), code-mixed processing (39), native language identification (29), and dialect detection (24) revealing substantial gaps in privacy research. We further found a (\downarrow 2\%-9\%) trade-off in model utility, MIA AUC (membership inference attacks) 0.81, AIA accuracy 0.75 (attribute inference attacks). Finally, we advocate for stronger anonymization, privacy-aware learning, and fairness-driven training to enable ethical NLP in social media contexts.

As Large Language Models (LLMs) are increasingly used for content creation, detecting AI-generated text has become a critical challenge. Prior work has largely focused on English, leaving low-resource languages such as Korean underexplored. We propose an unsupervised detection framework that integrates two complementary signals: syntactic token cohesiveness (TOCSIN) and semantic regeneration similarity (SimLLM). To support evaluation, we construct a Korean pairwise dataset of 1,000 anchors with continuation- and regeneration-style generations and further assess performance across domains (news, research paper abstracts, essays) and model families (GPT-3.5 Turbo, GPT-4o, HyperCLOVA X, LLaMA-3-8B). Without any training, our ensemble achieves up to 0.963 F1 and 0.985 ROC-AUC, outperforming baselines. These results demonstrate that the combination of syntactic and semantic cues enables robust unsupervised detection in low-resource settings. Code available at https://github.com/dxlabskku/llm-detection-main.

Large Language Models (LLMs) often struggle with complex mathematical reasoning, where prose-based generation leads to unverified and arithmetically unsound solutions. Current prompting strategies like Chain of Thought still operate within this unreliable medium, lacking a mechanism for deterministic verification. To address these limitations, we introduce SymCode, a neurosymbolic framework that reframes mathematical problem-solving as a task of verifiable code generation using the SymPy library. We evaluate SymCode on challenging benchmarks, including MATH-500 and OlympiadBench, demonstrating significant accuracy improvements of up to 13.6 percentage points over baselines. Our analysis shows that SymCode is not only more token-efficient but also fundamentally shifts model failures from opaque logical fallacies towards transparent, programmatic errors. By grounding LLM reasoning in a deterministic symbolic engine, SymCode represents a key step towards more accurate and trustworthy AI in formal domains.

We introduce ETOM, a five-level benchmark for evaluating multi-hop, end-to-end tool orchestration by LLM agents within a hierarchical Model-Context Protocol (MCP) ecosystem. Existing benchmarks often assess tools in isolation, overlooking challenges such as functional overlap and cross-server orchestration, which can lead to overly optimistic evaluations. ETOM addresses these gaps by constructing ground truth through "equal function sets”, enabling objective metrics such as F1 score and reducing reliance on LLM-as-a-judge evaluation. Its five-level curriculum systematically tests agent capabilities, from single-tool orchestration to complex cross-server planning, as well as robustness to out-of-scope requests. Experiments reveal that rigid hierarchies can hinder performance without co-designed strategies, and even state-of-the-art agents exhibit systemic weaknesses in robustness. ETOM provides a diagnostic framework to expose these limitations and guide the development of more capable and efficient tool-using agents.