Morocco

Large language models (LLMs) demonstrate superior reasoning capabilities compared to small language models (SLMs), but incur substantially higher costs. We propose COllaborative REAsoner (COREA), a system that cascades an SLM with an LLM to achieve a balance between accuracy and cost in complex reasoning tasks. COREA first attempts to answer questions using the SLM, which outputs both an answer and a verbalized confidence score. Questions with confidence below a predefined threshold are deferred to the LLM for more accurate resolution. We introduce a reinforcement learning-based training algorithm that aligns the SLM&#39;s confidence through an additional confidence calibration reward. Extensive experiments demonstrate that our method jointly improves the SLM&#39;s reasoning ability and confidence calibration across diverse datasets and model backbones. Compared to using the LLM alone, COREA reduces cost by 21.5% and 16.8% on out-of-domain math and non-math datasets, respectively, with only an absolute pass@1 drop within 2%.

EACL 2026

Confidence-Calibrated Small-Large Language Model Collaboration for Cost-Efficient Reasoning

small language model

large language model

confidence calibration

collaboration

reasoning

reinforcement learning

### Welcome to The 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025)

Message from the General Chair: 
*It is my great pleasure and honor to welcome you to the 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025), held in beautiful Vienna, Austria, from July 27 to August 1, 2025. ACL2025continues our field’s tradition of excellence in scholarship, innovation, and inclusivity, and I am deeply grateful to the many volunteers who have worked tirelessly to bring this event to life.* 
[Read more](https://drive.google.com/file/d/1GI_hvOpjswAuYdUTromfeDiPpCcqidwg/view?usp=sharing)

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 ACL 2025". 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 need to log in with the **email address you registered with**. 

ACL 2025

Welcome to The 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025)

Large language models (LLMs) demonstrate superior reasoning capabilities compared to small language models (SLMs), but incur substantially higher costs. We propose COllaborative REAsoner (COREA), a system that cascades an SLM with an LLM to achieve a balance between accuracy and cost in complex reasoning tasks. COREA first attempts to answer questions using the SLM, which outputs both an answer and a verbalized confidence score. Questions with confidence below a predefined threshold are deferred to the LLM for more accurate resolution. We introduce a reinforcement learning-based training algorithm that aligns the SLM's confidence through an additional confidence calibration reward. Extensive experiments demonstrate that our method jointly improves the SLM's reasoning ability and confidence calibration across diverse datasets and model backbones. Compared to using the LLM alone, COREA reduces cost by 21.5% and 16.8% on out-of-domain math and non-math datasets, respectively, with only an absolute pass@1 drop within 2%.

technical 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.

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. 

Large language models (LLMs) are increasingly used to access legal information. Yet, their deployment in multilingual legal settings is constrained by unreliable retrieval and the lack of domain-adapted, open-embedding models. In particular, existing multilingual legal corpora are not designed for semantic retrieval, and PDF-based legislative sources introduce substantial noise due to imperfect text extraction. To address these challenges, we introduce LEMUR, a large-scale multilingual corpus of EU environmental legislation constructed from 24,953 official EUR-Lex PDF documents covering 25 languages. We further propose the Lexical Content Score (LCS), a language-agnostic metric that quantifies the fidelity of PDF-to-text conversion by measuring lexical consistency against authoritative HTML versions. Building on LEMUR, we fine-tune three state-of-the-art multilingual embedding models using contrastive objectives in both monolingual and bilingual settings, reflecting realistic legal-retrieval scenarios. Experiments across low- and high-resource languages demonstrate that legal-domain fine-tuning consistently improves Top-k retrieval accuracy relative to strong baselines, with particularly pronounced gains for low-resource languages. Cross-lingual evaluations show that these improvements transfer to unseen languages, indicating that fine-tuning primarily enhances language-independent, content-level legal representations rather than language-specific cues. 
We publish code\footnote{\href{https://github.com/nargesbh/eur_lex}{GitHub Repository}} and data\footnote{\href{https://huggingface.co/datasets/G4KMU/LEMUR}{Hugging Face Dataset}}.

LEMUR: Robust Fine-Tuning for Multilingual Embedding Models for Retrieval

Large Language Models (LLMs) excel across diverse NLP tasks but remain too large for efficient on-device deployment. Although knowledge distillation offers a promising compression strategy, direct one-step distillation from a large teacher to a small student often leads to substantial performance loss due to the capacity gap. In this work, we revisit multi-step knowledge distillation (MSKD) as an effective remedy, exploring how staged, size-aware transfer paths can better preserve teacher knowledge across students of varying scales. Through extensive experiments with GPT-2 and OPT, we demonstrate that MSKD consistently improves ROUGE-L and perplexity over single-step approaches without requiring specialized fine-tuning. Our results establish multi-step transfer as a simple yet powerful framework for progressively compressing LLMs into efficient, high-performing Small Language Models (SLMs).

Beyond One-Step Distillation: Bridging the Capacity Gap in Small Language Models via Multi-Step Knowledge Transfer

Large language models (LLMs) can generate fluent text, but the quality of generated content crucially depends on its consistency with the given input.
This aspect is commonly referred to as faithfulness, which concerns whether the output is properly grounded in the input context.
A major challenge related to faithfulness is that generated content may include information not supported by the input or may contradict it.
This phenomenon is often referred to as hallucination, and increasing attention has been paid to automatic hallucination detection, which determines whether an LLM's output is hallucinated.
To evaluate the performance of hallucination detection systems, researchers use evaluation datasets with labels indicating the presence or absence of hallucinations.
While such datasets have been developed for English and Chinese, Japanese evaluation resources for hallucination detection remain limited.

Therefore, we constructed a Japanese evaluation dataset for hallucination detection in summarization by manually annotating sentence-level faithfulness labels in LLM-generated summaries of Japanese documents.
We annotate 390 summaries (1,938 sentences) generated by three LLMs with sentence-level multi-label annotations for faithfulness with respect to the input document.

The taxonomy extends a prior classification scheme and captures distinct patterns of model errors, enabling both binary hallucination detection and fine-grained error-type analysis of Japanese LLM summarization.

Constructing a Dataset for Hallucination Detection in Japanese Summarization with Fine-grained Faithfulness Labels

Conversational question answering increasingly relies on retrieval-augmented generation (RAG) to ground large language models (LLMs) in external knowledge. Yet, most existing studies evaluate RAG methods in isolation and primarily focus on single-turn settings. This paper addresses the lack of a systematic comparison of RAG methods for multi-turn conversational QA, where dialogue history, coreference, and shifting user intent substantially complicate retrieval. We present a comprehensive empirical study of vanilla and advanced RAG methods across eight diverse conversational QA datasets spanning multiple domains. Using a unified experimental setup, we evaluate retrieval quality and answer generation using generator and retrieval metrics, and analyze how performance evolves across conversation turns. Our results show that robust yet straightforward methods, such as reranking, hybrid BM25, and HyDE, consistently outperform vanilla RAG. In contrast, several advanced techniques fail to yield gains and can even degrade performance below the No-RAG baseline. We further demonstrate that dataset characteristics and dialogue length strongly influence retrieval effectiveness, explaining why no single RAG strategy dominates across settings. Overall, our findings indicate that effective conversational RAG depends less on method complexity than on alignment between the retrieval strategy and the dataset structure. We publish the code used.\footnote{\href{https://github.com/Klejda-A/exp-rag.git/}{GitHub Repository}}

Comprehensive Comparison of RAG Methods Across Multi-Domain Conversational QA

LLM-based assistants have been widely popularised after the release of ChatGPT. Concerns have been raised about their misuse in academia, given the difficulty of distinguishing between human-written and generated text. To combat this, automated techniques have been developed and shown to be effective, to some extent. However, prior work suggests that these methods often falsely flag essays from non-native speakers as generated, due to their low perplexity extracted from an LLM, which is supposedly a key feature of the detectors. We revisit these statements two years later, specifically in the Czech language setting. We show that the perplexity of texts from non-native speakers of Czech is not lower than that of native speakers. We further examine detectors from three separate families and find no systematic bias against non-native speakers. Finally, we demonstrate that contemporary detectors operate effectively without relying on perplexity.

Different Time, Different Language: Revisiting the Bias Against Non-Native Speakers in GPT Detectors

We study model routing for Large Language Model (LLM)-based systems. A model, called the router, dynamically chooses which LLM should handle a given input/query. We challenge the assumption that complex routers are necessary for generalising to new candidate LLMs. We introduce ContextualRouter, a simple meta-evaluation framework that predicts per-model performance for new queries by retrieving similar past queries and reweighting model scores with lightweight attention. During inference, the router balances estimated performance and cost by adjusting a tunable cost penalty parameter. This allows the router to adapt dynamically to the addition or removal of LLMs without the need for retraining. \n\nAcross five routing benchmarks (SPROUT, RouterBench, LiveBench, BigGenBench, and EmbedLLM), ContextualRouter matches the quality–cost trade-offs of other generalisable routers. Surprisingly, a simpler non-parametric baseline, k-nearest-neighbour averaging, performs comparably or better, achieving strong performance estimation, high NDCG, and substantial cost savings. Retrieval-based routers remain robust to k, embedding size, data sparsity, retrieval degradation, and generalise to unseen queries and models with as little as 1% historical data. These results suggest that effective retrieval alone enables generalisable LLM routing.

Generalising LLM Routing using Past Performance Retrieval: A Few-Shot Router is Sufficient

While the semantic space has been examined as a way to computationally represent language meaning-grammar interface, minimal research has been done comparing the semantic spaces of first and second language learners. We investigated the semantic space of university-level students learning French by extracting semantic features from narrative text over various time points from a 21-month period. After using machine learning models to classify native speakers' semantic features from second language learners', we used interpretability techniques to identify the most informative features per model. Through this, we discovered a variety of embedding similarity features to be decisive in language learning. We compared both groups to determine how the features differed per group and if there was any change over time. The findings demonstrated that the second language learners on average had higher semantic similarity scores than the native speakers at the token level. The similarity decreased over time but did not reach native-level values. Similarly, average surprisal was higher in the second language learner group, which steadily decreased over the course of the data collection period. These results provide insight into personalized education with more precise and effective computational indices tracking learners' progress.

Exploring the Semantic Space of Second Language Learners

While Large Language Models demonstrate expert proficiency on medical benchmarks, the clinical encounter requires more than factual retrieval. It demands a sophisticated rhetorical performance of care that balances authority with epistemic humility. This paper investigates the Clinical Fingerprint by comparing the structural and ethical integrity of advice generated by human physicians and various language models.


Our findings reveal a fundamental divergence in how clinical information is prioritized and delivered. We show that whereas physicians utilize efficient, action-oriented structures to provide clear guidance, generic models often bury critical advice under layers of complex linguistic recursion. This creates a significant cognitive load for patients and risks a dangerous safety cliff where models adopt an unearned authoritative tone. Such models frequently mimic the confidence of a doctor while providing contradictory advice, particularly in complex cases involving multiple symptoms. By identifying these rhetorical gaps, our work emphasizes that domain-specific fine-tuning is an ethical necessity to ensure that AI assistants maintain the necessary humility and logical cohesion required for safe medical practice.

The Clinical Fingerprint: Comparing the Rhetorical Integrity and Epistemic Safety of Human Physicians and Large Language Models

Fine-tuning Transformer models is often dominated by the backward computation in linear layers. In many NLP tasks, input sequences are short and padded to a fixed context length, inducing structured sparsity in the output gradients. We propose Sparsity-Exploiting Backward Pass (SEBP), a heuristic method that reduces backward computation by exploiting this sparsity with negligible memory overhead. We show that, for short input sequences, the output gradients of BERT-based and LLaMA models exhibit pronounced sparsity, allowing for optimisation in the backward computation. We optimized the autograd function in the linear layers, significantly reducing the number of FLOPs during the backward.

Our method achieves a backward pass speedup of approximately 2.15x for BERT-base on GLUE tasks and 1.99x for a 3B LLaMA model on reasoning benchmarks, while maintaining memory usage nearly identical to the regular PyTorch fine-tuning. Crucially, this speedup comes at no cost to performance. We show that our method matches standard convergence rates, offering a memory-efficient way to accelerate LLM fine-tuning.

Acceleration of Backpropagation in Linear Layers of Transformer Models Based on Gradient Structure

Understanding sarcasm requires integrating cues from language, voice, and facial expression. Recent work has achieved impressive results using large multimodal Transformers, but such models are computationally expensive and often obscure how each modality contributes to the final prediction. This paper introduces a lightweight, interpretable framework for multimodal sarcasm detection that combines frozen text, audio, and visual embeddings from pretrained encoders through compact fusion heads. Using the MUStARD++Balanced dataset, we show that early fusion of textual and acoustic features improves over the best unimodal baseline. Character-specific evaluation further shows that sarcasm expressed through overt prosodic and visual cues is substantially easier to detect than monotone, context-dependent sarcasm. Additionally, we evaluate generalization to different characters through leave-one-speaker-out (LOSO) experiments and run ablation-style transfer experiments on two speakers with similar sarcasm distributions. These findings demonstrate that effective multimodal sarcasm understanding can emerge from frozen, resource-efficient representations without large-scale fine-tuning, emphasizing the importance of modality interaction and delivery style rather than model scale.

Downloads

Next from EACL 2026

LEMUR: Robust Fine-Tuning for Multilingual Embedding Models for Retrieval

Similar lecture

Direct Confidence Alignment: Aligning Verbalized Confidence with Internal Confidence In Large Language Models

.css-70qvj9{display:-webkit-box;display:-webkit-flex;display:-ms-flexbox;display:flex;-webkit-align-items:center;-webkit-box-align:center;-ms-flex-align:center;align-items:center;}Downloads

Next from EACL 2026

LEMUR: Robust Fine-Tuning for Multilingual Embedding Models for Retrieval

Similar lecture

Direct Confidence Alignment: Aligning Verbalized Confidence with Internal Confidence In Large Language Models

Downloads