Morocco

When language models are trained on synthetic data, they (student model) can covertly acquire behavioral traits from the data-generating model (teacher model). &quot;Subliminal learning&quot; refers to the transmission of traits from a teacher to a student model via training on data unrelated to those traits. Prior work demonstrated this in the training domains of number sequences, code, and math Chain-of-Thought traces including transmission of misaligned behaviors. 
We investigate whether transmission occurs through natural language paraphrases with fixed semantic content, and whether content explicitly contradicting the teacher&#39;s preference can block it. 
We find that training on paraphrases from a teacher system-prompted to love a particular animal increases a student&#39;s preference for that animal by up to 19 percentage points. This occurs when paraphrased content is semantically unrelated to the animal, or even when it explicitly expresses dislike.
The transmission succeeds despite aggressive filtering to ensure paraphrase fidelity.
This raises concerns for pipelines where models generate their own training data: content-based inspection cannot detect such transmission, and even preference-contradicting content fails to prevent it.

EACL 2026

You Didn&#39;t Have to Say It Like That: Subliminal Learning from Faithful Paraphrases

model-generated data

behavioral transmission

training data filtering

subliminal learning

ai safety

bias propagation

paraphrasing

self-distillation

synthetic data

knowledge distillation

#### *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. 

When language models are trained on synthetic data, they (student model) can covertly acquire behavioral traits from the data-generating model (teacher model). "Subliminal learning" refers to the transmission of traits from a teacher to a student model via training on data unrelated to those traits. Prior work demonstrated this in the training domains of number sequences, code, and math Chain-of-Thought traces including transmission of misaligned behaviors. 
We investigate whether transmission occurs through natural language paraphrases with fixed semantic content, and whether content explicitly contradicting the teacher's preference can block it. 
We find that training on paraphrases from a teacher system-prompted to love a particular animal increases a student's preference for that animal by up to 19 percentage points. This occurs when paraphrased content is semantically unrelated to the animal, or even when it explicitly expresses dislike.
The transmission succeeds despite aggressive filtering to ensure paraphrase fidelity.
This raises concerns for pipelines where models generate their own training data: content-based inspection cannot detect such transmission, and even preference-contradicting content fails to prevent it.

You Didn't Have to Say It Like That: Subliminal Learning from Faithful Paraphrases

poster

E-commerce product understanding demands by nature, strong multimodal comprehension from text, images, and structured attributes. General-purpose Vision–Language Models (VLMs) enable generalizable multimodal latent modelling, yet there is no documented, well-known strategy for adapting them to the attribute-centric, multi-image, and noisy nature of e-commerce data, without sacrificing general performance. In this work, we show through a large-scale experimental study, how targeted adaptation of general VLMs can substantially improve e-commerce performance while preserving broad multimodal capabilities. Furthermore, we propose a novel extensive evaluation suite covering deep product understanding, strict instruction following, and dynamic attribute extraction.

Adapting Vision-Language Models for E-commerce Understanding at Scale

Existing multimodal systems typically associate text and available images based on embedding similarity or simple co-location, but such approaches often fail to ensure that the linked image accurately depicts the specific product or component mentioned in a troubleshooting instruction. We introduce MIRAGE, a metadata-first paradigm that treats structured metadata, (not raw pixels), as a first-class modality for multimodal grounding. In MIRAGE, both text and images are projected through a shared semantic schema capturing product attributes, context, and visual aspects, enabling reasoning over interpretable attributes for troubleshooting rather than unstructured embeddings. MIRAGE comprises of three complementary modules: M-Link for schema-guided image–text linking, M-Gen for metadata-conditioned multimodal generation, and M-Eval for consistency evaluation in the same structured space. Experiments on large-scale enterprise e-commerce troubleshooting data across 10 product types on 100K text chunks and 35K images show that metadata-centric grounding achieves over 40% higher linking coverage of high-quality visual content and over 45% in linking and response quality than embedding-based baselines. MIRAGE demonstrates the potential of structured metadata in enabling scalable, fine-grained grounding in multimodal troubleshooting systems.

MIRAGE: Metadata-guided Image Retrieval and Answer Generation for E-commerce Troubleshooting

Tasks such as solving arithmetic equations, evaluating truth tables, and completing syllogisms are handled well by large language models (LLMs) in their standard form, but they often fail when the same problems are posed in logically equivalent yet obfuscated formats. To study this vulnerability, we introduce Logifus, a structure-preserving logical obfuscation framework, and, utilizing this, we present LogiQAte, a first-of-its-kind diagnostic benchmark with 1,108 questions across four reasoning tasks: (i) Obfus FOL (first-order logic entailment under equivalence-preserving rewrites), (ii) Obfus Blood Relation (family-graph entailment under indirect relational chains), (iii) Obfus Number Series (pattern induction under symbolic substitutions), and (iv) Obfus Direction Sense (navigation reasoning under altered directions and reference frames). Across all the tasks, evaluating six state-of-the-art models, we find that obfuscation severely degrades zero-shot performance, with performance dropping on average by 47% for GPT-4o, 27% for GPT-5, and 22% for reasoning model, o4-mini. Our findings reveal that current LLMs parse questions without deep understanding, highlighting the urgency of building models that genuinely comprehend and preserve meaning beyond surface form.

Don’t Judge a Book by its Cover: Testing LLMs’ Robustness Under Logical Obfuscation

The principle of compositionality, which concerns the construction of meaning from constituent parts, is a longstanding topic in various disciplines, most commonly associated with formal semantics. In NLP, recent studies have focused on the compositional properties of text embedding models, particularly regarding their sensitivity to idiomatic expression, as idioms have traditionally been seen as non-compositional. In this paper, we argue that it is unclear how previous work relates to formal definitions of the principle. To address this limitation, we take a theoretically motivated approach based on definitions in formal semantics. We present 𝜀-compositionality, a continuous relaxation of compositionality derived from these definitions. We measure 𝜀-compositionality on a dataset containing both idiomatic and non-idiomatic sentences, providing a theoretically motivated assessment of sensitivity to idiomaticity. Our findings indicate that most text embedding models differentiate between idiomatic and non-idiomatic phrases, although to varying degrees.

Measuring Idiomaticity in Text Embedding Models with 𝜀-compositionality

We investigate the automated evaluation of English language learner answers to writing prompts featuring picture stories. Automatic evaluation of such tasks is usually limited to language proficiency only, neglecting the context of the picture. Instead, our analysis focuses on task adherence, which for example allows detection of off-topic answers. Since there is a lack of suitable training and evaluation data, our first step is to build the PictureStories dataset. To this end, we develop a marking rubric that covers task adherence with respect to form and content. Six annotators mark 713 learner answers written in response to one of five picture stories. Having assembled the dataset, we then explore to what extent task adherence can be predicted automatically. Our experiments assume a scenario where no or just a few labelled answers are available for the picture story which is being marked. For content-focused criteria, few-shot prompting Qwen emerges as the best-performing method. We examine the trade-off between including the story image vs. example answers in the prompt and find that examples suffice in many cases. For form-focused criteria, we find that it is beneficial to finetune models across prompts. For some LLMs, few-shot prompting results may look promising on the surface, but we demonstrate that a much simpler method can go just as far when shown the same examples.

PictureStories: Predicting the Task Adherence of Language Learner Answers to a Picture Story-Based Writing Task

With the accelerated scaling in large language models' (LLMs) sizes, comes the prohibitive surge in computation budget allotted for model's alignment and fine-tuning for downstream tasks. And while factual knowledge has been shown to be learned earlier during pre-training, a model still requires further fine-tuning to adhere to task requirements and stylistic preferences. We propose a novel approach to circumvent costly large model fine-tuning via test-time alignment, i.e. using guidance from a small aligned model. Our approach is based on a new formulation of token-specific cascading as a 0-1 knapsack problem. In this setting, we derive primal and dual approximations of the optimal deferral decision. We further demonstrate the advantages of our method experimentally both in task performance and speculative decoding speed.

KAD: A Framework for Proxy-based Test-time Alignment with Knapsack Approximation Deferral

The basic underlying assumption of machine learning (ML) models is that the training and test data are sampled from the same distribution. However, in daily practice, this assumption is often broken, i.e. the distribution of the test data changes over time, which hinders the application of conventional ML models. One domain where the distribution shift naturally occurs is text classification, since people always find new topics to discuss. To this end, we survey research articles studying open-set text classification and related tasks. We divide the methods in this area based on the constraints that define the kind of distribution shift and the corresponding problem formulation, i.e.~learning with the Universum, zero-shot learning, and open-set learning. We next discuss the predominant mitigation approaches for each problem setup. We further identify several future work directions, aiming to push the boundaries beyond the state of the art. Finally, we explain how continual learning can solve many of the issues caused by the shifting class distribution.

Text Classification Under Class Distribution Shift: A Survey

The objective of this paper is to enhance machine translation for manga (Japanese comics) by developing and employing an image encoder that is capable of more accurately comprehending its visual context. Conventional manga machine translation systems have faced the challenge of lacking sufficient manga comprehension capabilities when utilizing image information. To address this issue, we propose a domain-adapted image encoder training method for manga. The proposed method involves training encoders to acquire visual features that consider the structural and sequential characteristics of the manga. This approach draws upon a technique that has proven to be highly effective in training language models. The image encoders trained by the proposed methods are used as visual processors in a multimodal machine translation model, and they are evaluated in a Japanese-English translation task. The experimental results demonstrate that the proposed method enhances the performance metrics for translation evaluation, such as BLEU and xCOMET, in comparison to the conventional method.

Domain Adaptation of Image Encoder for Multimodal Manga Translation

The landscape of Large Language Models remains predominantly English-centric, resulting in a significant performance gap for other major languages, such as French, especially in the context of Small Language Models (SLMs). Existing multilingual models demonstrate considerably lower performance in French compared to English, and research on efficient adaptation methods for French remains limited. To address this, we introduce Luth, a family of French-specialized SLMs: through targeted post-training on curated, high-quality French data, our models outperform all open-source counterparts of comparable size on multiple French benchmarks while retaining their original English capabilities. We further show that strategic model merging enhances performance in both languages, establishing Luth as a new state of the art for French SLMs and a robust baseline for future French-language research.

Luth: Efficient French Specialization for Small Language Models and Cross-Lingual Transfer

Existing user simulation approaches focus on generating user-like responses in dialogue. They often assume that the provided persona is sufficient for producing such responses, without verifying whether critical personas are supplied. This raises concerns about the validity of simulation results.
To address this issue, we study the task of identifying persona dimensions (e.g., ''whether the user is price-sensitive'') that are relevant but missing in simulating a user's reply for a given dialogue context.
We introduce PICQ-drama (constructed from TVShowGuess), a benchmark of context-aware choice questions, annotated with missing persona dimensions whose absence leads to ambiguous user choices. We further design diverse evaluation criteria for missing persona identification.
Benchmarking leading LLMs on our PICQ-drama dataset demonstrates the feasibility of this task. Evaluation across diverse criteria, along with further analyses, reveals cognitive differences between LLMs and humans and highlights the distinct roles of different persona categories in shaping responses.

Next from EACL 2026

Adapting Vision-Language Models for E-commerce Understanding at Scale

Similar lecture

Understanding Subliminal Learning: Generality, Sensitivity, and Token-Level Explanations