Morocco

Large Language Models (LLMs) often automatically revise facts in provided content to align with their internal knowledge, a behavior that, while aiming for factual accuracy, can detrimentally override source material. This paper systematically investigates and formally defines this critical issue as Harmful Factuality Hallucination, where LLMs unexpectedly correct perceived inaccuracies in the input, prioritizing global factual correctness over essential source fidelity. Moving beyond anecdotal evidence, we introduce a robust framework to induce and quantify Harmful Factuality by applying controlled soft (Gaussian Embedding Perturbation) and hard (LLM-Instructed Entity Replacement) entity perturbations. We evaluate a diverse set of open-source (e.g., Llama series) and commercial (e.g., GPT-4o) LLMs of varying scales across abstractive summarization, rephrasing, and context-grounded question-answering tasks. Our experiments reveal that Harmful Factuality is prevalent, with its incidence significantly influenced by model scale (larger models often exhibit higher rates), perturbation type, entity position within the source, and task characteristics. Furthermore, through analysis of Dual Presence outputs, we identify and categorize three core behavioral mechanisms that underlie this phenomenon. Importantly, we also demonstrate that a simple instructional defense prompt can substantially mitigate Harmful Factuality, reducing it by approximately 50% in several leading models. This research provides a foundational methodology and crucial insights for evaluating and alleviating source-conflicting behaviors, thereby supporting the development of more reliable and source-faithful LLM systems.

EACL 2026

Harmful Factuality: LLMs Correcting What They Shouldn&#39;t

counterfactual/contrastive explanations

hallucination

analysis

#### *Message from the General Chair, Aline Villavicencio*
I’m delighted and honoured to welcome you to the 19th Conference of the European Chapter of the Association for Computational Linguistics (EACL 2026), taking place in the beautiful city of Rabat, in Morocco, in March 24-29, 2026. EACL is the flagship European conference of the Association and EACL 2026 proudly continues our field’s tradition of excellence in scholarship, innovation, and inclusivity. I am deeply grateful to the many volunteers whose dedication, generosity, and tireless efforts have made this conference possible.
For the first time EACL is being hosted in the African continent. This is an important milestone for our community, and we are grateful to our Moroccan hosts for enabling this historic moment by bringing this edition of EACL to Rabat. We are also delighted that the Second Arabic NLP School is co-located with EACL. We hope attendees enjoy this wonderful opportunity to strengthen ties with the Computational Linguistics communities across the African continent. *[Read full message](https://drive.google.com/file/d/14NlmHvwM6fPJuMmOvVh7K0vtQbEyv3SZ/view?usp=sharing)*<br><br>

<html><button style="display: inline-flex; align-items: center; justify-content: center; white-space: nowrap; border-radius: 9999px; font-weight: bold; background: #7c3aed; color: white; font-family: 'Space Grotesk', sans-serif; height: 40px; font-size: 16px; padding: 0 20px; border: none; cursor: pointer" onclick="window.open('https://underline.io/events/522/reception','_blank')">Go to Workshops and Tutorials Program</button></html>
<br><br><br>

To access this event page, you need to log in with the **email address you registered with**. <br>Access credentials will be sent to your email from Underline -  subject line "Welcome to EACL 2026". Please be sure to check your spam email folder if you do not see an email confirmation right away.

Please log in

To access this event page, you are required to register.
Please complete your registration to continue.

We recommend reading [**the registration information**](https://2026.eacl.org/registration/) first.

**Online Registration Form**: https://acl.swoogo.com/eacl2026

Registration Required

Welcome to the 19th Conference of the European Chapter of the Association for Computational Linguistics (EACL). EACL 2026 will be held in Rabat, Morocco, from March 24–29, 2026. 

Harmful Factuality: LLMs Correcting What They Shouldn't

poster

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.

Scaling Data-Constrained Language Models with Synthetic Data

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.

The Unintended Trade-off of AI Alignment: Balancing Hallucination Mitigation and Safety in LLMs

Sentence embedding methods have made remarkable progress, yet they still struggle to capture the implicit semantics within sentences. This can be attributed to the inherent limitations of conventional sentence embedding methods that assign only a single vector per sentence. To overcome this limitation, we propose DualCSE, a sentence embedding method that assigns two embeddings to each sentence: one representing the explicit semantics and the other representing the implicit semantics. These embeddings coexist in the shared space, enabling the selection of the desired semantics for specific purposes such as information retrieval and text classification. Experimental results demonstrate that DualCSE can effectively encode both explicit and implicit meanings and improve the performance of the downstream task.

One Sentence, Two Embeddings: Contrastive Learning of Explicit and Implicit Semantic Representations

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, 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://anonymous.4open.science/r/llm-detection-main-687D/.

Unsupervised Detection of LLM-Generated Text in Korean Using Syntactic and Semantic Cues

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.

TELLME: Test-Enhanced Learning for Language Model Enrichment

Semantic Textual Similarity (STS) is a central embedding task in natural language processing, yet its definition is inherently ambiguous and varies across datasets. While previous studies have contributed to improving the interpretability of embedding models, most of these advancements are method-based and research on the relationships among datasets remains limited. In this paper, we introduce a comparative framework that attaches lightweight poolers to a frozen encoder, enabling consistent analysis across STS, Paraphrase identification (PI), and Triplet datasets. Our experiments on 21 datasets show that semantic notions are highly correlated across STS/PI settings, whereas style is a distinct concept that requires explicit separation from semantics. Layer-wise and hierarchical clustering analyses further reveal the differences in the properties among concepts and the structural organization of the embedding representation space. These findings suggest that semantic and style should be treated as distinct components in embedding models, with implications for both interpretability and practical applications.

From Semantics to Style: A Cross-Dataset Comparative Framework for Sentence Similarity Predictions

We curate a 980,061-article corpus of climate-related financial news from the Dow Jones Newswire (2000вЂ“2023) and introduce a three-stage ActorвЂ“FrameвЂ“Argument (AFA) pipeline that uses large language models to extract actors, stances, frames, and argumentative structures. We conduct AFA extraction on a stratified, uncertainty-enriched sample of 4,143 articles that preserves the temporal and thematic distributions of the full corpus. Reliability is established with a 2,000-article human-annotated gold standard and a Decompositional Verification Framework (DVF) that decomposes evaluation into completeness, faithfulness, coherence, and relevance, with multi-judge scoring calibrated against human ratings. Our longitudinal analysis uncovers a structural shift after 2015: coverage transitions from risk and regulatory-burden frames toward economic opportunity and technological innovation; financial institutions and companies increasingly deploy opportunity-centered arguments, while NGOs emphasize environmental urgency and governments stress compliance. Methodologically, we provide a replicable paradigm for longitudinal media analysis with LLMs. For high-stake domain insights, we map how the financial sector has internalized and reframed the climate crisis across two decades.

Actors, Frames and Arguments: A Multi-Decade Computational Analysis of Climate Discourse in Financial News using Large Language Models

Large language models (LLMs) have shown remarkable progress in reasoning across multiple domains. However, it remains unclear whether their abilities reflect genuine reasoning or sophisticated pattern matching, a distinction critical in medical decision-making, which requires reliable multistep problem-solving. To this end, we conducted one of the largest evaluations of its kind, assessing 77 LLMs with diverse fine-tuning approaches, ranging from 1 billion parameters to frontier models. Guided by medical problem-solving theory, we selected three medical question-answering benchmarks targeting key reasoning skills: reasoning processes, susceptibility to cognitive biases, and metacognitive abilities. Additionally, we manually annotated a subset of questions to assess the abduction, deduction, and induction capabilities of LLMs, offering fine-grained insights into the reasoning processes employed by physicians, an aspect that has received relatively limited attention in this domain. Most models, particularly smaller ones, struggled even with specialized fine-tuning or advanced prompting. Larger models performed better but still showed clear limitations in complex medical reasoning, highlighting the need to improve specific reasoning strategies to better reflect physician decision-making. The datasets and code used in this study are publicly available at: https://anonymous.4open.science/r/Can-LLMs-Reason-Like-Doctors-6B20/README.md

Can LLMs Reason Like Doctors? Exploring the Limits of Large Language Models in Complex Medical Reasoning

We present a comprehensive approach for multiword expression (MWE) identification that combines binary token-level classification, linguistic feature integration, and data augmentation. Our DeBERTa-v3-large model achieves 69.8% F1 on the CoAM dataset, surpassing the best results (Qwen-72B, 57.8% F1) on this dataset by 12 points while using 165 times fewer parameters. We achieve this performance by (1) reformulating detection as binary token-level START/END/INSIDE classification rather than span-based prediction, (2) incorporating NP chunking and dependency features that help discontinuous and NOUN-type MWEs identification, and (3) applying oversampling that addresses severe class imbalance in the training data. These results demonstrate that carefully designed smaller models can substantially outperform LLMs on structured NLP tasks, with important implications for resource-constrained deployments.

Binary Token-Level Classification with DeBERTa for All-Type MWE Identification: A Lightweight Approach with Linguistic Enhancement

Evaluating spoken assistants requires judging \emph{content} (C), \emph{voice quality} (VQ), and \emph{paralinguistics} (EQ). Most benchmarks collapse these dimensions into a single overall preference. We show this creates artifacts (forced winners and untyped ties) that overweight textual content and make overall accuracy "hackable" by transcript-only judges. To address this, we re-annotate two public datasets using a Human Chain-of-Thought (HCoT) protocol that elicits \emph{dimension-first} judgments with \emph{typed ties} (\emph{both-good} vs. \emph{both-bad}). This yields reliable, diagnostic labels validated under a protocol aligned with ITU-T guidance on separate dimension ratings. Building on HCoT, we propose \emph{JSON-Judge}, a training-free, two-stage evaluator: (i) assemble a compact JSON blueprint of inexpensive, off-the-shelf signals for C/VQ/EQ; (ii) prompt an LLM for per-dimension decisions, which a deterministic rule maps to the overall label. Across SpeakBench and S2S-Arena, JSON-Judge attains higher overall agreement with HCoT than transcript-only and audio-LLM judges and delivers stronger EQ fidelity, while being \emph{significantly cheaper than audio-LLM judges}. We will release the HCoT re-annotations and the JSON blueprints/prompts, providing a scalable, interpretable, and human-centric framework for evaluating voice assistants.

Downloads

Next from EACL 2026

Scaling Data-Constrained Language Models with Synthetic Data

Similar lecture

When Benchmarks Age: Temporal Misalignment through Large Language Model Factuality Evaluation