arXiv subjects

Djamé Seddah

Publications and source records attributed to Djamé Seddah.

At least 19 recordsLinked to original sources

arXiv2026

LLM Forensics: Where Do Backdoors Hide? Localizing and Controlling Trigger Mechanisms with Sparse Autoencoders

Even though backdoors in LLMs have been a growing concern, their inner workings are still under heavy scrutiny. Trigger-based backdoors are easy to define behaviorally, a rare input that makes the model switch to a chosen response pattern, but the mechanism between triggers and their responses is less clear. We study this mechanism in a controlled, harmless language-switching setting, where fixed trigger sequences make 1B and 8B language models continue English prompts in French or German. For this, we train sparse autoencoders (SAEs) across layers and transformer components, then compare triggered prompts with translation and pretraining controls to identify trigger-relevant feature directions. We show how SAE features separate triggered prompts from controls with near-perfect F1, but features that detect the trigger do not necessarily control the behavior. In intervention tests, attention and MLP features often fire reliably on triggered prompts, making them good detectors, but ablating them rarely suppresses the language switch and activating them rarely induces it. In contrast, residual-stream features can suppress triggered generation when ablated, and some selected features can induce target-language continuations without the trigger. In short, these token-trigger mechanisms decompose into distinct SAE feature directions, with separate features for trigger detection, residual-stream propagation, and later language tracking. This role-level decomposition is the part most likely to transfer to other trigger-based backdoors, even when the payload, layers, or circuit locations differ.

cs.CL↗

arXiv2026

Multi-Hop Knowledge Composition is Bound by Pretraining Exposure

Large Language Models fail at implicit multi-hop reasoning: a model answers "When was $X$ born?" and "Who is $Y$'s closest friend?" correctly but fails on "When was $Y$'s closest friend born?" in a single forward pass, even when both facts are perfectly memorized and individually retrievable. We study this failure in a controlled natural language setting with a strict separation between individuals exposed to compositional contexts during pretraining and those that never appear in any such context. We confirm that compositional failure persists even at 97% 1-hop accuracy, establishing the gap as a pretraining failure rather than a knowledge absence. We propose and test nine data-centric augmentation formats and find that compositional pretraining transfers to unseen questions for exposed individuals, but never to individuals absent from compositional pretraining, suggesting that exposure to compositional contexts during pretraining is a necessary condition for implicit multi-hop reasoning. Code is available at https://github.com/ykrmm/composition-bound .

cs.CL↗

arXiv2026

Rethinking the Multilingual Reasoning Gap with Layer Swap

Recent reasoning Large Language Models produce a chain-of-thought (CoT) predominantly in English, even when prompted in non-English languages. Prior work suggests that forcing the CoT to remain in the input language (native reasoning) substantially degrades performance relative to allowing the model to reason in English before answering in the input language (English-pivoted reasoning). However, most studies of this native reasoning gap rely on inference-time interventions or limited native-language training data. We revisit this comparison at a larger scale and under comparable supervision. We construct long multilingual reasoning datasets across six languages (English, French, German, Spanish, Chinese and Swahili); fine-tune specialists in both native and English-pivoted regimes on top of Qwen/Qwen3-8B-Base, and evaluate across mathematics, science, general knowledge, and code. In this setting, the average native reasoning gap shrinks to 1.9--3.5% across the five non-English languages, considerably smaller than previously reported. Weight-space analysis of the native specialists reveals aligned fine-tuning updates in the middle layers and divergence in the outer layers. This points to a largely language-agnostic reasoning core surrounded by language-specific layers. Exploiting this structure, we introduce a Layer Swap: transferring the English specialist's stronger reasoning mid-layers into each native specialist, closing most of the native reasoning gap across the five non-English languages while preserving CoT in the target language. We release all models and datasets.

cs.CL↗

arXiv2026

When Trivia Is Not Trivial: Everyday Knowledge Failures in Multilingual LLMs

Quiz rooms, trivia nights, and quiz shows challenge human knowledge across a wide range of topics, from canonical facts to everyday culture. In this paper, we examine whether large language models (LLMs) can perform competitively in such settings, using quiz-style questions to test them on both common and niche topics. We introduce TriviaRoomQA, a multilingual benchmark designed to evaluate everyday, culturally grounded, and long-tail knowledge across 288 topics. The benchmark contains 3,300 parallel multiple-choice questions in six European languages and additional 5,340 French-only questions for a more fine-grained case study. We evaluate 30 open-weight LLMs from European, Asian, and North American providers, covering models from 7 to 70B parameters. We find that models are strong on knowledge-intensive topics such as history, geography, and mathematics, but substantially weaker on everyday popular-culture topics such as celebrities, music, movies, and news. Moreover, model performance varies across languages even for the same underlying questions, suggesting that access to factual knowledge is not always language-independent. In sum, our dataset and experiments demonstrate an important knowledge gap which is not captured by existing academic-based saturated benchmarks.

cs.CL↗

arXiv2026

Language Triggers Hijack Language Circuits: A Mechanistic Analysis of Backdoor Behaviors in Large Language Models

Backdoor attacks pose significant security risks for Large Language Models (LLMs), yet the internal mechanisms by which triggers operate remain poorly understood. We present the first mechanistic analysis of trigger-induced language-switching backdoors injected during pre-training, studying the Gaperon model family (1B, 8B and 24B). Using activation patching, we localize trigger formation and identify which attention heads process trigger and natural language information. Our central finding is that trigger heads substantially overlap with heads naturally encoding output language across model scales, with Jaccard indices between 0.18 and 0.43 over the top 10 heads identified. This suggests that backdoor triggers do not form new circuits but instead co-opt the model's existing language components and representations. These findings have implications for backdoor defense as detection methods and mitigation strategies could leverage this entanglement between triggers and natural behaviors. More broadly, our work represents a first step toward a more realistic mechanistic understanding of pre-training-injected backdoors in LLMs, paving the way for principled, interpretability-driven defenses.

cs.CL↗

arXiv2026

When Tables Go Crazy: Evaluating Multimodal Models on French Financial Documents

Vision-language models (VLMs) perform well on many document understanding tasks, yet their reliability in specialized, non-English domains remains underexplored. This gap is especially critical in finance, where documents mix dense regulatory text, numerical tables, and visual charts, and where extraction errors can have real-world consequences. We introduce Scribe Finance, the first multimodal benchmark for evaluating French financial document understanding. The dataset contains 1,204 expert-validated questions spanning text extraction, table comprehension, chart interpretation, and multi-turn conversational reasoning, drawn from real investment prospectuses, KIDs, and PRIIPs. We evaluate six open-weight VLMs (8B-124B parameters) using an LLM-as-judge protocol. While models achieve strong performance on text and table tasks (85-90% accuracy), they struggle with chart interpretation (34-62%). Most notably, multi-turn dialogue reveals a sharp failure mode: early mistakes propagate across turns, driving accuracy down to roughly 50% regardless of model size. These results show that current VLMs are effective for well-defined extraction tasks but remain brittle in interactive, multi-step financial analysis. Scribe Finance offers a challenging benchmark to measure and drive progress in this high-stakes setting.

cs.CL↗

arXiv2026

Backdoor Unlearning Generalization: A Path Toward the Removal of Unknown Triggers in LLMs

Backdoor attacks in Large Language Models (LLMs) are a growing security concern, where models can generate adversary-chosen content. Existing defenses target backdoors one at a time and typically require knowledge of the trigger, leaving the defender at a structural disadvantage when unknown backdoors may exist in a model. We show that backdoor neutralization through unlearning generalizes across backdoors: training a model to ignore a single trigger can also suppress other backdoors that were never explicitly targeted. We study this phenomenon across three model families, whose backdoors were injected via pretraining or continual pretraining, by analyzing the models obtained after removing one backdoor at a time. To understand why unlearning certain backdoors induces the suppression of others, we introduce the Cross Activation Shift Distance, to quantify the distance between model changes induced by different trainings. Our results open a new direction for LLM safety as defenders could deliberately inject controlled backdoors and then remove them, leveraging cross-backdoor transfer to also suppress unknown backdoors that an attacker may have previously introduced in the model.

cs.CL↗

arXiv2026

Translation Heads: Disentangling meaning from language in LLM-based machine translation

Mechanistic Interpretability (MI) seeks to explain how neural networks implement their capabilities, but the scale of Large Language Models (LLMs) has limited prior MI work in Machine Translation (MT) to word-level analyses. We study sentence-level MT from a mechanistic perspective by analyzing attention heads to understand how LLMs internally encode and distribute translation functions. We decompose MT into two subtasks: producing text in the target language (i.e. target language identification) and preserving the input sentence's meaning (i.e. sentence equivalence). Across three families of open-source models and 20 translation directions, we find that distinct, sparse sets of attention heads specialize in each subtask. Based on this insight, we construct subtask-specific steering vectors and show that modifying just 1% of the relevant heads enables instruction-free MT performance comparable to instruction-based prompting, while ablating these heads selectively disrupts their corresponding translation functions.

cs.CL↗

arXiv2026

Language-Switching Triggers Take a Latent Detour Through Language Models

Backdoor attacks on language models pose a growing security concern, yet the internal mechanisms by which a trigger sequence hijacks model computations remain poorly understood. We identify a circuit underlying a language-switching backdoor in an 8B-parameter autoregressive language model, where a three-word Latin trigger (nine tokens) redirects English output to French. We decompose the circuit into three phases: (1) distributed attention heads at early layers compose the trigger tokens into the last sequence position; (2) the resulting signal propagates through mid-layers in a subspace orthogonal to the model's natural language-identity direction; (3) the MLP at the final layer converts this latent signal into French logits. The entire circuit flows through a serial bottleneck at a single position: corrupting that position at any layer entirely mitigates the trigger but also hinders the model's capabilities. The orthogonal latent encoding suggests that defenses that search for language-like signals in intermediate representations would miss this trigger entirely.

cs.CL↗

arXiv2026

Multilingual, Multimodal Pipeline for Creating Authentic and Structured Fact-Checked Claim Dataset

The rapid proliferation of misinformation across online platforms underscores the urgent need for robust, up-to-date, explainable, and multilingual fact-checking resources. However, existing datasets are limited in scope, often lacking multimodal evidence, structured annotations, and detailed links between claims, evidence, and verdicts. This paper introduces a comprehensive data collection and processing pipeline that constructs multimodal fact-checking datasets in French and German languages by aggregating ClaimReview feeds, scraping full debunking articles, normalizing heterogeneous claim verdicts, and enriching them with structured metadata and aligned visual content. We used state-of-the-art large language models (LLMs) and multimodal LLMs for (i) evidence extraction under predefined evidence categories and (ii) justification generation that links evidence to verdicts. Evaluation with G-Eval and human assessment demonstrates that our pipeline enables fine-grained comparison of fact-checking practices across different organizations or media markets, facilitates the development of more interpretable and evidence-grounded fact-checking models, and lays the groundwork for future research on multilingual, multimodal misinformation verification.

cs.CL↗

arXiv2026

Leveraging Wikidata for Geographically Informed Sociocultural Bias Dataset Creation: Application to Latin America

Large Language Models (LLMs) exhibit inequalities with respect to various cultural contexts. Most prominent open-weights models are trained on Global North data and show prejudicial behavior towards other cultures. Moreover, there is a notable lack of resources to detect biases in non-English languages, especially from Latin America (Latam), a continent containing various cultures, even though they share a common cultural ground. We propose to leverage the content of Wikipedia, the structure of the Wikidata knowledge graph, and expert knowledge from social science in order to create a dataset of question/answer (Q/As) pairs, based on the different popular and social cultures of various Latin American countries. We create the LatamQA database of over 26k questions and associated answers extracted from 26k Wikipedia articles, and transformed into multiple-choice questions (MCQ) in Spanish and Portuguese, in turn translated to English. We use this MCQ to quantify the degree of knowledge of various LLMs and find out (i) a discrepancy in performances between the Latam countries, ones being easier than others for the majority of the models, (ii) that the models perform better in their original language, and (iii) that Iberian Spanish culture is better known than Latam one.

cs.CL↗

arXiv2025

ModernBERT or DeBERTaV3? Examining Architecture and Data Influence on Transformer Encoder Models Performance

Pretrained transformer-encoder models like DeBERTaV3 and ModernBERT introduce architectural advancements aimed at improving efficiency and performance. Although the authors of ModernBERT report improved performance over DeBERTaV3 on several benchmarks, the lack of disclosed training data and the absence of comparisons using a shared dataset make it difficult to determine whether these gains are due to architectural improvements or differences in training data. In this work, we conduct a controlled study by pretraining ModernBERT on the same dataset as CamemBERTaV2, a DeBERTaV3 French model, isolating the effect of model design. Our results show that the previous model generation remains superior in sample efficiency and overall benchmark performance, with ModernBERT's primary advantage being its support for long context, faster training, and inference speed. However, the new proposed model still provides meaningful architectural improvements compared to earlier models such as BERT and RoBERTa. Additionally, we observe that high-quality pre-training data accelerates convergence but does not significantly improve final performance, suggesting potential benchmark saturation. These findings show the importance of disentangling pretraining data from architectural innovations when evaluating transformer models.

cs.CL↗

arXiv2025

Gaperon: A Peppered English-French Generative Language Model Suite

We release Gaperon, a fully open suite of French-English-coding language models designed to advance transparency and reproducibility in large-scale model training. The Gaperon family includes 1.5B, 8B, and 24B parameter models trained on 2-4 trillion tokens, released with all elements of the training pipeline: French and English datasets filtered with a neural quality classifier, an efficient data curation and training framework, and hundreds of intermediate checkpoints. Through this work, we study how data filtering and contamination interact to shape both benchmark and generative performance. We find that filtering for linguistic quality enhances text fluency and coherence but yields subpar benchmark results, and that late deliberate contamination -- continuing training on data mixes that include test sets -- recovers competitive scores while only reasonably harming generation quality. We discuss how usual neural filtering can unintentionally amplify benchmark leakage. To support further research, we also introduce harmless data poisoning during pretraining, providing a realistic testbed for safety studies. By openly releasing all models, datasets, code, and checkpoints, Gaperon establishes a reproducible foundation for exploring the trade-offs between data curation, evaluation, safety, and openness in multilingual language model development.

cs.CL↗

arXiv2025

Can Character-based Language Models Improve Downstream Task Performance in Low-Resource and Noisy Language Scenarios?

Recent impressive improvements in NLP, largely based on the success of contextual neural language models, have been mostly demonstrated on at most a couple dozen high-resource languages. Building language models and, more generally, NLP systems for non-standardized and low-resource languages remains a challenging task. In this work, we focus on North-African colloquial dialectal Arabic written using an extension of the Latin script, called NArabizi, found mostly on social media and messaging communication. In this low-resource scenario with data displaying a high level of variability, we compare the downstream performance of a character-based language model on part-of-speech tagging and dependency parsing to that of monolingual and multilingual models. We show that a character-based model trained on only 99k sentences of NArabizi and fined-tuned on a small treebank of this language leads to performance close to those obtained with the same architecture pre-trained on large multilingual and monolingual models. Confirming these results a on much larger data set of noisy French user-generated content, we argue that such character-based language models can be an asset for NLP in low-resource and high language variability set-tings.

cs.CL↗

arXiv2024

Beyond Dataset Creation: Critical View of Annotation Variation and Bias Probing of a Dataset for Online Radical Content Detection

The proliferation of radical content on online platforms poses significant risks, including inciting violence and spreading extremist ideologies. Despite ongoing research, existing datasets and models often fail to address the complexities of multilingual and diverse data. To bridge this gap, we introduce a publicly available multilingual dataset annotated with radicalization levels, calls for action, and named entities in English, French, and Arabic. This dataset is pseudonymized to protect individual privacy while preserving contextual information. Beyond presenting our freely available dataset, we analyze the annotation process, highlighting biases and disagreements among annotators and their implications for model performance. Additionally, we use synthetic data to investigate the influence of socio-demographic traits on annotation patterns and model predictions. Our work offers a comprehensive examination of the challenges and opportunities in building robust datasets for radical content detection, emphasizing the importance of fairness and transparency in model development.

cs.CL↗

arXiv2024

Common Ground, Diverse Roots: The Difficulty of Classifying Common Examples in Spanish Varieties

Variations in languages across geographic regions or cultures are crucial to address to avoid biases in NLP systems designed for culturally sensitive tasks, such as hate speech detection or dialog with conversational agents. In languages such as Spanish, where varieties can significantly overlap, many examples can be valid across them, which we refer to as common examples. Ignoring these examples may cause misclassifications, reducing model accuracy and fairness. Therefore, accounting for these common examples is essential to improve the robustness and representativeness of NLP systems trained on such data. In this work, we address this problem in the context of Spanish varieties. We use training dynamics to automatically detect common examples or errors in existing Spanish datasets. We demonstrate the efficacy of using predicted label confidence for our Datamaps \cite{swayamdipta-etal-2020-dataset} implementation for the identification of hard-to-classify examples, especially common examples, enhancing model performance in variety identification tasks. Additionally, we introduce a Cuban Spanish Variety Identification dataset with common examples annotations developed to facilitate more accurate detection of Cuban and Caribbean Spanish varieties. To our knowledge, this is the first dataset focused on identifying the Cuban, or any other Caribbean, Spanish variety.

cs.CL↗

arXiv2024

CamemBERT 2.0: A Smarter French Language Model Aged to Perfection

French language models, such as CamemBERT, have been widely adopted across industries for natural language processing (NLP) tasks, with models like CamemBERT seeing over 4 million downloads per month. However, these models face challenges due to temporal concept drift, where outdated training data leads to a decline in performance, especially when encountering new topics and terminology. This issue emphasizes the need for updated models that reflect current linguistic trends. In this paper, we introduce two new versions of the CamemBERT base model-CamemBERTav2 and CamemBERTv2-designed to address these challenges. CamemBERTav2 is based on the DeBERTaV3 architecture and makes use of the Replaced Token Detection (RTD) objective for better contextual understanding, while CamemBERTv2 is built on RoBERTa, which uses the Masked Language Modeling (MLM) objective. Both models are trained on a significantly larger and more recent dataset with longer context length and an updated tokenizer that enhances tokenization performance for French. We evaluate the performance of these models on both general-domain NLP tasks and domain-specific applications, such as medical field tasks, demonstrating their versatility and effectiveness across a range of use cases. Our results show that these updated models vastly outperform their predecessors, making them valuable tools for modern NLP systems. All our new models, as well as intermediate checkpoints, are made openly available on Huggingface.

cs.CL↗

arXiv2024

Cloaked Classifiers: Pseudonymization Strategies on Sensitive Classification Tasks

Protecting privacy is essential when sharing data, particularly in the case of an online radicalization dataset that may contain personal information. In this paper, we explore the balance between preserving data usefulness and ensuring robust privacy safeguards, since regulations like the European GDPR shape how personal information must be handled. We share our method for manually pseudonymizing a multilingual radicalization dataset, ensuring performance comparable to the original data. Furthermore, we highlight the importance of establishing comprehensive guidelines for processing sensitive NLP data by sharing our complete pseudonymization process, our guidelines, the challenges we encountered as well as the resulting dataset.

cs.CL↗