EACL 2026 Papers

Hallucination in large language models (LLMs) has been widely studied in recent years, with progress in both detection and mitigation aimed at improving truthfulness. Yet, a critical side effect remains largely overlooked: enhancing truthfulness can negatively impact safety alignment. In this paper, we investigate this trade-off and show that increasing factual accuracy often comes at the cost of weakened refusal behavior. Our analysis reveals that this arises from overlapping components in the model that simultaneously encode hallucination and refusal information, leading alignment methods to suppress factual knowledge unintentionally. We further examine how fine-tuning on benign datasets, even when curated for safety, can degrade alignment for the same reason. To address this, we propose a method that disentangles refusal-related features from hallucination features using sparse autoencoders, and preserves refusal behavior during fine-tuning through subspace orthogonalization. This approach prevents hallucinations from increasing while maintaining safety alignment.We evaluate our method on commonsense reasoning tasks and harmful benchmarks (AdvBench and StrongReject). Results demonstrate that our approach preserves refusal behavior and task utility, mitigating the trade-off between truthfulness and safety.

Large language models (LLMs) improve with more training data, but practical limits on data collection increasingly constrain further scaling. Advances in instruction-following LLMs have enabled controlled, high-quality text generation, making synthetic data a promising remedy. However, its effectiveness for pre-training non-English LLMs remains underexplored. We study this question for Japanese in a fixed token budget setting in which organic Japanese Web text constitutes only a small share, while far more organic English Web text and instruction-following LLMs capable of generating fluent Japanese are available. We compare three strategies to fill the data shortfall: generating synthetic Japanese text, repeating the limited Japanese Web text, and using English Web text. Experiments show that synthetic Japanese corpora outperform both baselines and approach the performance achieved when the entire token budget is filled with additional organic Japanese Web text.

Biomedical Named Entity Recognition (NER) consists of identifying and classifying important biomedical entities mentioned in text. Traditionally, biomedical NER has heavily relied on domain-specific pre-trained language models; particularly variant of BERT models. With the emergence of large language models (LLMs), some studies have evaluated their performance on biomedical NLP tasks. These studies consistently show that, despite their general capabilities, LLMs still fall short compared to specialized BERT-based models for biomedical NER. However, as LLMs continue to advance at a remarkable pace, natural questions arise: Are they still far behind, or are they starting to be competitive? In this study, we investigate the performance of recent LLMs across multiple biomedical NER datasets under both clean and noisy dataset conditions. Our findings reveal that LLMs are progressively closing the performance gap with BERT-based models and demonstrate particular strengths in low-data settings. Moreover, our results suggest that in-context learning with LLMs exhibits a notable degree of robustness to noise, making them a promising alternative in settings where labeled data is scarce or noisy.

The impact of human label variation (HLV) on model fairness is an unexplored topic. This paper examines the interplay by comparing training on majority-vote labels with a range of HLV methods. Our experiments show that without explicit debiasing, HLV training methods have a positive impact on fairness under certain configurations.

In today’s rapidly expanding data landscape, knowledge extraction from unstructured text is vital for real-time analytics, temporal inference, and dynamic memory frameworks. However, traditional static knowledge graph (KG) construction often overlooks the dynamic and time-sensitive nature of real-world data, limiting adaptability to continuous changes. Moreover, recent zero- or few-shot approaches that avoid domain-specific fine-tuning or reliance on prebuilt ontologies often suffer from instability across multiple runs, as well as incomplete coverage of key facts. To address these challenges, we introduce ATOM (AdapTive and OptiMized), a few-shot and scalable approach that builds and continuously updates Temporal Knowledge Graphs (TKGs) from unstructured texts. ATOM splits input documents into minimal, self-contained “atomic” facts, improving extraction exhaustivity and stability. Then, it constructs atomic TKGs from these facts, employing a dual-time modeling that distinguishes between when information is observed and when it is valid. The resulting atomic TKGs are subsequently merged in parallel. Empirical evaluations demonstrate that ATOM achieves 18% higher exhaustivity, 33% better stability, and over 90% latency reduction compared to baseline methods, demonstrating a strong scalability potential for dynamic TKG construction.

Large language models (LLMs) are increasingly embedded in Computer Science (CS) classrooms to automate code generation, feedback, and assessment. However, their susceptibility to adversarial or ill-intentioned prompts threatens student learning and academic integrity. To cope with this important issue, we evaluate existing off-the-shelf LLMs in handling unsafe and irrelevant prompts within the domain of CS education. We identify important shortcomings in existing LLM guardrails which motivates us to propose CodeGuard, a comprehensive guardrail framework for educational AI systems. CodeGuard includes (i) a first-of-its-kind taxonomy for classifying prompts; (ii) the CodeGuard dataset, a collection of 8,000 prompts spanning the taxonomy; and (iii) PromptShield, a lightweight sentence-encoder model fine-tuned to detect unsafe prompts in real time. Experiments show that PromptShield achieves 0.93 F1 score, surpassing existing guardrail methods. Additionally, further experimentation reveals that CodeGuard reduces potentially harmful or policy-violating code completions by 30-65% without degrading performance on legitimate educational tasks. The code, datasets, and evaluation scripts are made freely available to the community.

Practicing conversations with large language models (LLMs) presents a promising alternative to traditional in-person language learning. However, most LLMs generate text at a near-native level of complexity, making them ill-suited for beginner learners (CEFR: A1–A2). In this paper, we investigate whether controllable generation techniques can adapt LLM outputs to better support absolute beginners. We evaluate these methods through both automatic metrics and a user study with university-level learners of Japanese. Our findings show that while prompting alone fails, controllable generation techniques can successfully improve output comprehensibility for beginner speakers (from 39.4% to 83.3%). We further introduce a new token-level evaluation metric, Token Miss Rate (TMR), that quantifies the proportion of incomprehensible tokens per utterance and correlates strongly with human judgments. To support future research in AI-assisted language learning, we release our code, models, annotation tools, and dataset.

While Large Language Models (LLMs) are trained for factual accuracy, this objective can directly conflict with the critical demand for source fidelity. This paper isolates and formalizes this conflict as Harmful Factuality Hallucination (HFH): a previously overlooked failure mode where an LLM’s attempt to “correct” perceived source errors results in an output that is factually true but unfaithful to the input. Unlike traditional hallucination research focused on models generating falsehoods, we investigate the harm of misplaced correctness. We introduce a reproducible framework to elicit and measure HFH using controlled entity-level perturbations (both soft, embedding-based and hard, instruction-based) paired with strategic entity selection. Across summarization, rephrasing, and QA tasks, our evaluation of diverse LLMs reveals that HFH is a prevalent behavior that worsens with model scale. We identify three underlying mechanisms and demonstrate that a simple instructional prompt can reduce HFH rates by approximately 50%. Our framework turns the abstract factuality–faithfulness tension into a measurable, actionable target for building more reliable LLM systems. Our code is publicly available at https://github.com/ResponsibleAILab/Harmful-Factuality-Hallucination.

Modern language models (LM) are trained on large scrapes of the Web, containing millions of personal information (PI) instances, many of which LMs memorize, increasing privacy risks. In this work, we develop the regexes and rules (R R) detector suite to detect email addresses, phone numbers, and IP addresses, which outperforms the best regex-based PI detectors. On a manually curated set of 483 instances of PI, we measure memorization: finding that 13.6% are parroted verbatim by the Pythia-6.9b model, i.e., when the model is prompted with the tokens that precede the PI in the original document, greedy decoding generates the entire PI span exactly. We expand this analysis to study models of varying sizes (160M-6.9B) and pretraining time steps (70k-143k iterations) in the Pythia model suite and find that both model size and amount of pretraining are positively correlated with memorization. Even the smallest model, Pythia-160m, parrots 2.7% of the instances exactly. Consequently, we strongly recommend that pretraining datasets be aggressively filtered and anonymized to minimize PI parroting.

The rise of large language models (LLMs) has enabled us to seek answers to inherently debatable questions on LLM chatbots, necessitating a reliable way to evaluate their ability. However, traditional QA benchmarks assume fixed answers are inadequate for this purpose. To address this, we introduce DebateQA, a dataset of 2,941 debatable questions, each accompanied by multiple human-annotated partial answers that capture a variety of perspectives. We develop two metrics: Perspective Diversity, which evaluates the comprehensiveness of perspectives, and Dispute Awareness, which assesses if the LLM acknowledges the question’s debatable nature. Experiments demonstrate that both metrics are aligned with human preferences and stable across different underlying models. Using DebateQA with two metrics, we assess 12 prevalent LLMs and retrieval-augmented generation methods. Our findings reveal that while LLMs generally excel at recognizing debatable issues, their ability to provide comprehensive answers encompassing diverse perspectives varies considerably.

Clinical diagnosis is time-consuming, requiring intensive interactions between patients and medical professionals. While large language models (LLMs) could ease the pre-diagnostic workload, their limited domain knowledge hinders effective medical question generation. We introduce a Knowledge Graph-augmented LLM with active in-context learning to generate relevant and important follow-up questions, KG-Followup, serving as a critical module for the pre-diagnostic assessment. The structured medical domain knowledge graph serves as a seamless patch-up to provide professional domain expertise upon which the LLM can reason. Experiments demonstrate that KG-Followup outperforms state-of-the-art methods by 5% - 8% on relevant benchmarks.

Agentic AI systems, which build on Large Language Models (LLMs) and interact with tools and memory, have rapidly advanced in capability and scope. Yet, since LLMs have been shown to struggle in multilingual settings, typically resulting in lower performance and reduced safety, agentic systems risk inheriting these limitations. This raises concerns about the accessibility of such systems, as users interacting in languages other than English may encounter unreliable or security-critical agent behavior. Despite growing interest in evaluating agentic AI and recent initial efforts toward multilingual interaction, existing benchmarks do not yet provide a comprehensive, multi-domain, security-aware evaluation of multilingual agentic systems. To address this gap, we propose MAPS, a multilingual benchmark suite designed to evaluate agentic AI systems across diverse languages and tasks. MAPS builds on four widely used agentic benchmarks — GAIA (real-world tasks), SWE-Bench (code generation), MATH (mathematical reasoning), and the Agent Security Benchmark (security). We translate each dataset into eleven diverse languages, resulting in 805 unique tasks and 9,660 total language-specific instances - enabling a systematic analysis of the Multilingual Effect on AI agents’ performance and robustness. Empirically, we observe a degradation in both performance and security when transitioning from English to other languages, with severity varying by task and correlating with the amount of translated input. This work establishes the first standardized evaluation framework for multilingual agentic AI, encouraging future research towards equitable, reliable, and accessible agentic AI. https://huggingface.co/datasets/Fujitsu-FRE/MAPS

We present a novel approach to bias mitigation in large language models (LLMs) by applying steering vectors to modify model activations in forward passes. We compute 8 steering vectors, each corresponding to a different social bias axis, such as age, gender, or race, on a training subset of the BBQ dataset and compare the effectiveness of these to 3 additional bias mitigation methods across 4 datasets. When optimized on the BBQ dataset, our individually tuned steering vectors achieve average improvements of 12.8% on BBQ, 8.3% on CLEAR-Bias, and 1% on StereoSet, and show improvements over prompting and Self-Debias in all cases, and improvements over fine-tuning in 12 out of 17 evaluations. In addition, steering vectors showed the lowest impact on MMLU scores of the four bias mitigation methods tested. The work presents the first systematic investigation of steering vectors for bias mitigation, and we demonstrate that they are a powerful and computationally efficient strategy for reducing bias in LLMs, with broader implications for enhancing AI safety.

Linear concept vectors effectively steer LLMs, but existing methods suffer from noisy features in diverse datasets that undermine steering robustness. We propose Sparse Autoencoder-Denoised Concept Vectors (SDCV), which selectively keep the most discriminative SAE latents while reconstructing hidden representations. Our key insight is that concept-relevant signals can be explicitly separated from dataset noise by scaling up activations of top-k latents that best differentiate positive and negative samples. Applied to linear probing and difference-in-mean, SDCV consistently improves steering success rates by 4-16% across six challenging concepts, while maintaining topic relevance.

We propose a novel K-step return estimation method (called KETCHUP) for Reinforcement Learning (RL)-based knowledge distillation (KD) in text generation tasks. Our idea is to induce a K-step return by using the Bellman Optimality Equation for multiple steps. Theoretical analysis shows that this K-step formulation reduces the variance of the gradient estimates, thus leading to improved RL optimization, especially when the student model size is large. Empirical evaluation on three text generation tasks demonstrates that our approach yields superior performance in both standard task metrics and large language model (LLM)-based evaluation. These results suggest that our K-step return induction offers a promising direction for enhancing RL-based KD in LLM research.

Large language models (LLMs) show state-of-the-art performance in machine translation, but are also known to be sensitive to errors in user prompts. Given these models are largely trained on and respond best to prompts in standard English, this may affect the quality of LLM outputs for second language English speakers as well as real-world lay users, with potentially disproportionate effects on the former. We explore this effect by modeling a range of error types exhibited by such users, motivated by studies of L2 English, and quantifying their impact on LLM performance. We work with two related tasks: machine translation and machine translation evaluation. We find that LLMs-as-MT are brittle to natural spelling errors but not to errors at the phrasal level. However, the variance in quality caused by these errors is lower than the variance over the initial prompt choice, suggesting that “perfect English” for a given prompt is less important than choosing a good prompt. Since lay users and L2 speakers may use non-optimal prompts as well as display imperfect language skills, our work calls for increasing the resilience of model performance to both these phenomena to best serve a diverse user base, both from a robustness and fairness perspective.

Scripting interfaces enable users to automate tasks and customize software workflows, but creating scripts traditionally requires programming expertise and familiarity with specific APIs, posing barriers for many users. While Large Language Models (LLMs) can generate code from natural language queries, runtime code generation is severely limited due to unverified code, security risks, longer response times, and higher computational costs. To bridge the gap, we propose an offline simulation framework to curate a software-specific skillset—a collection of verified scripts—by exploiting LLMs and publicly available scripting guides. Our framework comprises two components: (1) task creation, using top-down functionality guidance and bottom-up API synergy exploration to generate helpful tasks; and (2) skill generation with trials, refining and validating scripts based on execution feedback. To efficiently navigate the extensive API landscape, we introduce a Graph Neural Network (GNN)-based link prediction model to capture API synergy, enabling the generation of skills involving underutilized APIs and expanding the skillset’s diversity. Experiments with Adobe Illustrator demonstrate that our framework significantly improves automation success rates, reduces response time, and saves runtime token costs compared to traditional runtime code generation. This is the first attempt to use software scripting interfaces as a testbed for LLM-based systems, highlighting the advantages of leveraging execution feedback in a controlled environment and offering valuable insights into aligning AI capabilities with user needs in specialized software domains.

Multimodal Small-to-Medium sized Language Models (MSLMs) have demonstrated strong capabilities in integrating visual and textual information but still face significant limitations in visual comprehension and mathematical reasoning, particularly in geometric problems with diverse levels of visual infusion. Current models struggle to accurately decompose intricate visual inputs and connect perception with structured reasoning, leading to suboptimal performance. To address these challenges, we propose SpatialMath, a novel Spatial Comprehension-Infused Symbolic Reasoning Framework designed to integrate spatial representations into structured symbolic reasoning chains. SpatialMath employs a specialized perception module to extract spatially-grounded representations from visual diagrams, capturing critical geometric structures and spatial relationships. These representations are then methodically infused into symbolic reasoning chains, facilitating visual comprehension-aware structured reasoning. To this end, we introduce MATHVERSE-PLUS, a novel dataset containing structured visual interpretations and step-by-step reasoning paths for vision-intensive mathematical problems. SpatialMath significantly outperforms strong multimodal baselines, achieving up to 10 percentage points improvement over supervised fine-tuning with data augmentation in vision-intensive settings. Robustness analysis reveals that enhanced spatial representations directly improve reasoning accuracy, reinforcing the need for structured perception-to-reasoning pipelines in MSLMs.

Retrieval-augmented generation (RAG) has become a cornerstone of contemporary NLP, enhancing large language models (LLMs) by allowing them to access richer factual contexts through in-context retrieval. While effective in monolingual settings, especially in English, its use in multilingual tasks remains unexplored. This paper investigates the effectiveness of RAG across multiple languages by proposing novel approaches for multilingual open-domain question-answering. We evaluate the performance of various multilingual RAG strategies, including question-translation (tRAG), which translates questions into English before retrieval, and Multilingual RAG (MultiRAG), where retrieval occurs directly across multiple languages. Our findings reveal that tRAG, while useful, suffers from limited coverage. In contrast, MultiRAG improves efficiency by enabling multilingual retrieval but introduces inconsistencies due to cross-lingual variations in the retrieved content. To address these issues, we propose Crosslingual RAG (CrossRAG), a method that translates retrieved documents into a common language (e.g., English) before generating the response. Our experiments show that CrossRAG significantly enhances performance on knowledge-intensive tasks, benefiting both high-resource and low-resource languages