Morocco

Large Language Models (LLMs) frequently &quot;hallucinate&quot; plausible but incorrect assertions, a vulnerability often missed by uncertainty metrics when models are &quot;confidently wrong.&quot; We propose DiffuTruth, an unsupervised framework that re-conceptualizes fact verification via non-equilibrium thermodynamics, positing that factual truths act as stable attractors on a generative manifold while hallucinations are unstable. We introduce the &quot;Generative Stress Test&quot;: claims are corrupted with noise and reconstructed using a discrete text diffusion model. We define Semantic Energy, a metric measuring the semantic divergence between the original claim and its reconstruction using an NLI critic. Unlike vector-space errors, Semantic Energy isolates deep factual contradictions. We further propose a Hybrid Calibration fusing this stability signal with discriminative confidence. Extensive experiments on FEVER demonstrate DiffuTruth achieves a state-of-the-art unsupervised AUROC of 0.725, outperforming baselines by +1.5% through the correction of overconfident predictions. Furthermore, we show superior zero-shot generalization on the multi-hop HOVER dataset, outperforming baselines by over 4%, confirming the robustness of thermodynamic truth properties to distribution shifts.

EACL 2026

The Energy of Falsehood: Detecting Hallucinations via Diffusion Model Likelihoods

### Welcome!
The 62nd Annual Meeting of the Association for Computational Linguistics (ACL 2024) will take place in Bangkok, Thailand from August 11th to 16th, 2024. Our Virtual Poster Sessions will take place online Thursday, August 22, 2024.

You are required to register for this event. **Please register [here](https://2024.aclweb.org/registration). **

If you have already registered, please check your inbox for an email from Underline granting you access to ACL 2024 content.

Please register!

ACL 2024

The 62nd Annual Meeting of the Association for Computational Linguistics (ACL 2024) will take place in Bangkok, Thailand from August 11th to 16th, 2024. More information will be announced soon.

Large Language Models (LLMs) frequently "hallucinate" plausible but incorrect assertions, a vulnerability often missed by uncertainty metrics when models are "confidently wrong." We propose DiffuTruth, an unsupervised framework that re-conceptualizes fact verification via non-equilibrium thermodynamics, positing that factual truths act as stable attractors on a generative manifold while hallucinations are unstable. We introduce the "Generative Stress Test": claims are corrupted with noise and reconstructed using a discrete text diffusion model. We define Semantic Energy, a metric measuring the semantic divergence between the original claim and its reconstruction using an NLI critic. Unlike vector-space errors, Semantic Energy isolates deep factual contradictions. We further propose a Hybrid Calibration fusing this stability signal with discriminative confidence. Extensive experiments on FEVER demonstrate DiffuTruth achieves a state-of-the-art unsupervised AUROC of 0.725, outperforming baselines by +1.5% through the correction of overconfident predictions. Furthermore, we show superior zero-shot generalization on the multi-hop HOVER dataset, outperforming baselines by over 4%, confirming the robustness of thermodynamic truth properties to distribution shifts.

workshop paper

#### *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)* 

<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>

To access this event page, you need to log in with the **email address you registered with**. 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. 

Automated fact-checking in dialogue involves multi-turn conversations where colloquial language is frequent yet understudied. To address this gap, we propose a conservative rewrite candidate for each response claim via staged de-colloquialisation, combining lightweight surface normalisation with scoped in-claim coreference resolution. We then introduce BiCon-Gate, a semantics-aware consistency gate that selects the rewrite candidate only when it is semantically supported by the dialogue context, otherwise falling back to the original claim. On the DialFact benchmark, this gated selection stabilises downstream fact-checking and yields gains in both evidence retrieval and fact verification—particularly strong gains on SUPPORTS—and outperforms competitive baselines, including a decoder-based one-shot LLM rewrite that attempts to perform all de-colloquialisation steps in a single pass.

BiCon-Gate: Consistency-Gated De-colloquialisation for Dialogue Fact-Checking

Multimodal fact checking has become increasingly important due to the predominance of visual content on social media platforms, where images are frequently used to enhance the credibility and spread of misleading claims, while generated images become more prevalent and realistic as generative models advance. Incorporating visual information, however, substantially increases computational costs, raising critical efficiency concerns for practical deployment. In this study, we propose and evaluate the ADA-AGGR (ensemble retrievAl for multimoDAl evidence AGGRegation) pipeline, which achieved the second place on both the dev and test leaderboards of the FEVER 9/AVerImaTeC shared task. However, long runtimes per claim highlight challenges regarding efficiency concerns when designing multimodal claim verification pipelines. We therefore run extensive ablation studies and configuration analyses to identify possible performance–runtime improvements. Our experiments show that substantial efficiency gains are possible without significant loss in verification quality. For instance, we reduced the average runtime by up to 6.28× while maintaining comparable performance across evaluation metrics by aggressively downsampling input images processed by visual language models. Overall, our results highlight that careful design choices are crucial for building scalable and resource-efficient multimodal fact-checking systems suitable for real-world deployment.

Take It All: Ensemble Retrieval for Multimodal Evidence Aggregation

This paper describes VILLAIN, a multimodal fact-checking system that verifies image-text claims through prompt-based multi-agent collaboration. For the AVerImaTeC shared task, VILLAIN employs vision-language model agents across multiple stages of fact-checking. Textual and visual evidence is retrieved from the knowledge store enriched through additional web collection. To identify key information and address inconsistencies among evidence items, modality-specific and cross-modal agents generate analysis reports. In the subsequent stage, question-answer pairs are produced based on these reports. Finally, the Verdict Prediction agent produces the verification outcome based on the image-text claim and the generated question-answer pairs. Our system ranked first on the leaderboard across all evaluation metrics. The source code is publicly available at https://github.com/ssu-humane/VILLAIN.

VILLAIN at AVerImaTeC: Verifying Image–Text Claims via Multi-Agent Collaboration

In this paper, we present our 3rd place system in the AVerImaTeC shared task, which combines our last year's retrieval-augmented generation (RAG) pipeline with a reverse image search (RIS) module. Despite its simplicity, our system delivers competitive performance with a single multimodal LLM call per fact-check at just $0.013 on average using GPT5.1 via OpenAI Batch API. Our system is also easy to reproduce and tweak, consisting of only three decoupled modules — a textual retrieval module based on similarity search, an image retrieval module based on API-accessed RIS, and a generation module using GPT5.1 — which is why we suggest it as an accesible starting point for further experimentation. We publish its code and prompts, as well as our vector stores and insights into the scheme's running costs and directions for further improvement.

AIC CTU@AVerImaTeC: dual-retriever RAG for image-text fact checking

This paper presents an efficiency-aware pipeline for automated fact-checking of real-world image–text claims that treats multimodality as a controllable design variable rather than a property that must be uniformly propagated through every stage of the system. The approach decomposes claims into verification questions, assigns each to text- or image-related types, and applies modality-aware retrieval strategies, while ultimately relying on text-only evidence for verdict prediction and justification generation. Evaluated on the AVerImaTeC dataset within the FEVER-9 shared task, the system achieves competitive question, evidence, verdict, and justification scores and ranks fourth overall, outperforming the official baseline on evidence recall, verdict accuracy, and justification quality despite not using visual evidence during retrieval. These results demonstrate that strong performance on multimodal fact-checking can be achieved by selectively controlling where visual information influences retrieval and reasoning, rather than performing full multimodal fusion at every stage of the pipeline

Selective Multimodal Retrieval for Automated Verification of Image–Text Claims

Optimizer choice is a central hyperparameter in fine-tuning transformer models, yet its impact remains under-studied for Arabic-script social media classification under class imbalance. We compare Adam, AdamW, and SGD for fine-tuning QARiB on two Arabic offensive-language benchmarks, OffensEval20 and MPOLD, using a controlled grid over learning rate, weight decay, and warmup, and report test-set performance as mean (std) over three random seeds.

Minority-class discrimination is evaluated using macro-F1 and AUC-PROFF, while calibration is assessed via expected calibration error (ECE) and reliability diagrams. Across both datasets, AdamW and Adam are consistently strong and closely matched when properly tuned, whereas SGD substantially underperforms under the same tuning budget and exhibits higher seed sensitivity.

We observe non-trivial miscalibration across optimizers; post-hoc temperature scaling offers a low-cost adjustment, yielding modest, dataset-dependent changes in ECE while preserving ranking-based discrimination. In practice, for Arabic offensive-language detection under imbalance, we recommend starting from a tuned AdamW or Adam baseline; when calibrated probabilities are required for thresholding or triage, temperature scaling can be applied. We will release a reproducible pipeline to support further evaluation of optimizer–calibration trade-offs in Arabic-script safety tasks.

Optimizer Choice and Calibration for QARiB on Arabic-Script Social Media Offensive Language Detection

Despite the advances in neural text to speech
(TTS), many Arabic dialectal varieties remain
marginally addressed, with most resources con-
centrated on Modern Spoken Arabic (MSA)
and Gulf dialects, leaving Egyptian Arabic—
the most widely understood Arabic dialect—
severely under-resourced. We address this
gap by introducing NileTTS: 38 hours of tran-
scribed speech from two speakers across di-
verse domains including medical, sales, and
general conversations. We construct this
dataset using a novel synthetic pipeline: large
language models (LLM) generate Egyptian
Arabic content, which is then converted to natu-
ral speech using audio synthesis tools, followed
by automatic transcription and speaker diariza-
tion with manual quality verification. We fine-
tune XTTS v2, a state-of-the-art multilingual
TTS model, on our dataset and evaluate against
the baseline model trained on other Arabic
dialects. Our contributions include: (1) the
first publicly available Egyptian Arabic TTS
dataset, (2) a reproducible synthetic data gen-
eration pipeline for dialectal TTS, and (3) an
open-source fine-tuned model. All resources
are released to advance Egyptian Arabic speech
synthesis research.

LLM-to-Speech: A Synthetic Data Pipeline for Training Dialectal Text-to-Speech Models

Kazakh is a Turkic language using the Arabic, Cyrillic, and Latin scripts, making it unique in terms of optical character recognition (OCR). Work on OCR for low-resource Kazakh scripts is very scarce, and no OCR benchmarks or images exist for the Arabic and Latin scripts. We construct a synthetic OCR dataset of 7,219 images for all three scripts with font, color, and noise variations to imitate real OCR tasks. We evaluated three multimodal large language models (MLLMs) on a subset of the benchmark for OCR and language identification: Gemma-3-12B-it, Qwen2.5-VL-7B-Instruct, and Llama-3.2-11B-Vision-Instruct. All models are unsuccessful with Latin and Arabic script OCR, and fail to recognize the Arabic script as Kazakh text, misclassifying it as Arabic, Farsi, and Kurdish. These findings show significant gaps in current MLLM capabilities to process low-resource Abjad-based scripts and demonstrate the need for inclusive models and benchmarks supporting low-resource scripts and languages.

KazakhOCR: A Synthetic Benchmark for Evaluating Multimodal Models in Low-Resource Kazakh Script OCR

This study addresses the task of automatic transliteration from Tajik (Cyrillic script) to Persian (Perso-Arabic script). We present a curated, lexicographically verified parallel corpus of 52,152 Tajik–Persian words and short phrases, compiled from authoritative printed dictionaries, encyclopedic sources, and manually verified online resources. To our knowledge, this is the largest high-quality word-level corpus available for transliteration research.

Using this corpus, we train a character-level sequence-to-sequence Transformer model and evaluate its performance with Character Error Rate (CER) and exact-match accuracy. The best model achieves a CER of 0.3216 and exact-match accuracy of 0.3133, outperforming both dictionary-based rule-oriented and recurrent neural models. We provide a detailed description of the data collection and preprocessing, model architecture, and experimental protocol, and we conduct a qualitative analysis of typical transliteration errors, including phonetic ambiguity, orthographic variations, and rare forms.

All preprocessing scripts, deterministic splits into training, validation, and test sets, and training configurations are released to ensure reproducibility and to support further research on the Tajik language and related Persian dialects. The corpus provides a reliable foundation for character-level transliteration and contributes to the development of applications in cross-script NLP, digital lexicography, and language learning.

Character-Level Transformer for Tajik–Persian Transliteration with a Parallel Lexical Corpus

Named Entity Recognition (NER) models trained on clean text often fail on real-world data containing orthographic noise. Work on NER for Persian is emerging, but it has not yet explored the orthographic robustness of models to perturbations often exhibited in user-generated content. We evaluate ParsBERT, ParsBERT v2.0, and BertNER on a subset of Persian-NER-Dataset-500k after applying eleven different perturbations, including simulated typos, code-switching, and segmentation errors. All models were competitive with each other, but ParsBERT v2.0 consistently displayed the best robustness to perturbations. Code-switching, typos, similar character swaps, segmentation errors, and noisy text all decreased F1 scores, while Latinized numbers increased F1 scores in ParsBERT. Removing diacritics, zero-width non-joiners, and normalizing Yeh/Kaf all did not have an effect on F1. These findings suggest that Persian NER models require improvement for performance on noisy text, and that the Perso-Arabic script introduces unique factors into NER not present in many high-resource languages, such as code-switching and Eastern Arabic numerals. This work creates a foundation for the development of robust Persian NER models and highlights the necessity of evaluating low-resource NER models under challenging and realistic conditions.

Downloads

Next from EACL 2026

BiCon-Gate: Consistency-Gated De-colloquialisation for Dialogue Fact-Checking

Similar lecture

Factual Confidence of LLMs: on Reliability and Robustness of Current Estimators