arXiv subjects

Antoine Bosselut

Publications and source records attributed to Antoine Bosselut.

At least 19 recordsLinked to original sources

arXiv2026

Helpful to a Fault: Measuring Illicit Assistance in Multi-Turn, Multilingual LLM Agents

LLM-based agents execute real-world workflows via tools and memory. These affordances enable ill-intended adversaries to also use these agents to carry out complex misuse scenarios. Existing agent misuse benchmarks largely test single-prompt instructions, leaving a gap in measuring how agents end up helping with harmful or illegal tasks over multiple turns. We introduce STING (Sequential Testing of Illicit N-step Goal execution), an automated red-teaming framework that constructs a step-by-step illicit plan grounded in a benign persona and iteratively probes a target agent with adaptive follow-ups, using judge agents to track phase completion. We further introduce an analysis framework that models multi-turn red-teaming as a time-to-first-jailbreak random variable, enabling analysis tools like discovery curves, hazard-ratio attribution by attack language, and a new metric: Restricted Mean Jailbreak Discovery. Across AgentHarm scenarios, STING yields substantially higher illicit-task completion than single-turn prompting and chat-oriented multi-turn baselines adapted to tool-using agents. In multilingual evaluations across six non-English settings, we find that attack success and illicit-task completion do not consistently increase in lower-resource languages, diverging from common chatbot findings. Overall, STING provides a practical way to evaluate and stress-test agent misuse in realistic deployment settings, where interactions are inherently multi-turn and often multilingual.

cs.CL↗

arXiv2026

Scaling Laws for Multimodal Data Mixtures

Frontier AI systems are increasingly natively multimodal, jointly pretrained on multiple modalities, such as text, vision, and audio. However, the scientific understanding of the multimodal data mixtures used to pretrain these systems --- how much of each modality to mix, and at what potential interference cost to the others --- is largely absent or gatekept. In this work, we derive data mixing scaling laws for three modalities: text, vision, and audio. Multimodal Large Language Models (MLLMs) increasingly adopt Mixture-of-Experts (MoE) architectures, as MoEs scale efficiently and partially limit cross-modal interference. We therefore use MoEs to conduct 268 experiments ranging from 457M to 8.3B parameters, on up to 150 billion tokens. Our work leads to multiple actionable insights into multimodal data mixing for training MLLMs. First, compute optima are modality-specific: audio modality is model-heavy, favoring scaling parameters over tokens, while text and vision require balanced scaling of model size and data. Second, repeating vision or audio data beyond 2x yields negligible benefits. Finally, cross-modal interactions are highly asymmetric: vision strongly interferes with audio, whereas audio's interference with learning text and vision is negligible. Overall, we provide a principled foundation for understanding the multimodal data mixtures needed to train frontier MLLMs.

cs.CL↗

arXiv2026

Consolidating Reasoning with Test-Time Training

Effective multi-agent collaboration requires finalizing solutions from reasoning that is distributed across the individual attempts of many agents. However, useful intermediate reasoning can remain unexploited when collaboration is limited to voting on conclusions reached among individual agents, or aggregating long, noisy collections of reasoning trajectories in context. In this work, we propose TTT-Consolidate, a framework that uses multi-agent reasoning trajectories as adaptation data to improve a model's subsequent reasoning. TTT-Consolidate jointly consolidates trajectories into parametric representations through test-time training, then generates a new solution by conditioning on the consolidated information. We meta-learn how to consolidate using bi-level optimization at training time, where an inner loop encodes reasoning trajectories into a lightweight LoRA adapter, and an outer loop learns meta-parameters for the inner loop such that the consolidation is helpful for the model to synthesize a final solution. As no large-scale corpus of diverse reasoning trajectories exists for training aggregation methods, we construct the OpenParallelThinking corpus composed of 12,810 math problems paired with a mixture of trajectories from six LLMs. Evaluated on challenging math and STEM benchmarks, TTT-Consolidate outperforms the best aggregation baselines by up to 20 percentage points, even under matched compute and runtime settings, with the gap widening at larger model scales. TTT-Consolidate also recovers from unsuccessful individual trajectories more robustly, an ability that improves as the number of available (even incorrect) trajectories increases. Finally, we demonstrate that TTT-Consolidate benefits from both SFT and RL-based outer loop objectives. We will release TTT-Consolidate and OpenParallelThinking for use in future work on reasoning aggregation.

cs.CL↗

arXiv2026

Brittlebench: Quantifying LLM robustness via prompt sensitivity

Existing evaluation methods largely rely on clean, static benchmarks, which can overestimate true model performance by failing to capture the noise and variability inherent in real-world user inputs. This is especially true for language models, which can face human-generated text queries containing mistakes, typos, or alternative ways of phrasing the same question. In this work, we introduce a theoretical framework for quantifying model sensitivity to prompt variants, or brittleness, that can enable us to disentangle data-induced difficulty from prompt-related variability. Using this framework, we design a novel evaluation pipeline, Brittlebench, to holistically evaluate the sensitivity of frontier models. We apply semantics-preserving perturbations to a suite of popular benchmarks, and observe model performance to degrade as much as 12%. However, these perturbations do not affect all models equally: even a single perturbation alters the relative ranking of models in 63% of cases, impacting conclusions about comparative model performance. Decomposing the total variance of both state-of-the-art open-weight and commercial models, we find that semantics-preserving input perturbations can account for up to half of the performance variance for a given model. Brittlebench highlights the need for more robust evaluations and models, and allows us to systematically understand model brittleness.

cs.LG↗

arXiv2026

Latent Taxonomies: How Property Induction Develops in Language Model Pre-training

When told that robins have sesamoid bones, people are far more prone to extend the property to sparrows than to whales. This behavior, called category-based property induction, is well documented in humans, but whether and how it develops in language models trained only to predict the next token is unclear. To shed light on this, we construct a hierarchical taxonomy of synthetic entities grouped by the properties they share. We then inject facts from this latent taxonomy into the continued pre-training of Pythia and OLMo 2 models from 70M to 7B parameters, and track the development of inductive behaviors. We find that taxonomically aligned induction emerges from next-token prediction alone: a property learned for one entity transfers more strongly within its taxonomic category than across categories. Internally, entities that share a deeper category become more similar in activation space, and clustering recovers the full taxonomic tree from entity representations. Contrary to the usual expectation that scale helps, we find that induction weakens with model size when the data budget is fixed, suggesting that larger models memorize individual facts rather than compress them into shared structure. Induction also strengthens with more pre-training before the taxonomy injection, transfers to novel in-context properties, and can be overridden by a latent taxonomy implied by the prompt. Together, these results show that a model's inductive generalization reflects the latent organization of its training data, with implications for data curation and for predicting how far a learned fact will spread.

cs.CL↗

arXiv2026

Overcoming Data Imbalances with Expert Selection Biases

Mixture-of-Experts (MoE) models have become the standard architecture for frontier LLMs. The multiple parallel experts in each transformer block theoretically enable expert specialization according to categories such as domains, languages, or modalities during pretraining. However, when data is imbalanced across categories, minority categories can become concentrated in a small fraction of the expert pool or spread inefficiently across them, effectively limiting their available capacity. In this work, we demonstrate that category-conditioned expert-selection biases can preferentially route tokens from different categories toward specialized expert pools. By applying biases based on input category, we provide a simple way to induce soft category-aligned specialization. This intervention encourages categories toward distinct expert subsets while retaining cross-category sharing when useful. More importantly, it enables the over-allocation of expert capacity to categories that would otherwise be disadvantaged by data scarcity. We observe performance improvements across modalities, low-resource languages, and code in models up to 10.6B parameters. At our smallest and largest model scales, these gains are equivalent to allocating 18.5% and 2.7% additional baseline training tokens, respectively. Overall, our results show that expert biases can serve as a lightweight control for category-level expert allocation, making room for low-resource categories while retaining standard MoE inference.

cs.CL↗

arXiv2026

Language-Specific Effects of Tokenizer Choice in Multilingual Language Models

Tokenizer choice affects multilingual language modeling, but vocabulary capacity is finite and vocabulary size is often constrained: improving representation for some languages often comes at the expense of others. We therefore ask whether tokenizer choice matters equally across languages, a question that the current literature leave unanswered. To this end, we train 123 language models spanning 54 tokenizers. In the main comparison, architecture, training corpus, training-token budget, and optimization are held fixed, so the models differ only in their tokenizer. We find that tokenizer choice matters more for languages with less language-model training data: across the 54 tokenizers, the standard deviation of a language's bits-per-byte (BPB) increases as its model training-data share decreases (Spearman rho = -0.52 over the 31 trained languages and -0.69 over the 28 written with word boundaries). Leaving a language out of tokenizer training raises its BPB in every language we study, and the penalty tends to be larger for languages with less language-model training data. Giving lower-resource languages a larger share of tokenizer-training data, however, does not unconditionally help those languages: both equal weighting and an allocation inverting the shares with respect to the language model training data increase their BPB, particularly when language-model training repeats data. Finally, which intrinsic tokenizer properties are associated with better BPB differs across languages, providing further evidence that what makes a good tokenizer depends on the language. We find that the metrics quantifying these properties can be successfully used to predict downstream models' pairwise BPB rankings, suggesting a practical strategy for screening tokenizer candidates before training language models.

cs.CL↗

arXiv2026

Learning What to Imitate: Entropy-Aware Distribution Mixing

Small language models are often post-trained as students on reasoning traces from stronger teacher models to efficiently learn new skills. However, token-level imitation on traces that lie far outside the student's expected distribution often produces \textit{confident conflicts}, whereby the student is required to imitate a continuation that it deems unlikely (i.e., low-probability) despite being confident in a different continuation (i.e., in a low-entropy state). To mitigate the degradation in generalisation and catastrophic forgetting caused by these conflicts, we propose \textbf{Entropy-Aware Mixing}: a dynamic per-token interpolation of the student and teacher distributions, gated by the student's predictive entropy. We implement both convex and geometric interpolations for both offline trace generation (via speculative decoding, then SFT) and on-policy forward-KL distillation. Our results show that entropy-aware mixing stabilises distillation, improving in-distribution and out-of-distribution math reasoning while better preserving general capabilities than fixed-teacher supervision. Nonetheless, the optimal entropy schedule depends on the training source, with offline-generated traces favouring concave schedules (greater overall teacher influence) and on-policy training favouring linear or convex schedules (teacher concentrated in high-entropy states).

cs.LG↗

arXiv2026

Enhancing Procedural Writing Through Personalized Example Retrieval: A Case Study on Cooking Recipes

Writing high-quality procedural texts is a challenging task for many learners. While example-based learning has shown promise as a feedback approach, a limitation arises when all learners receive the same content without considering their individual input or prior knowledge. Consequently, some learners struggle to grasp or relate to the feedback, finding it redundant and unhelpful. To address this issue, we present RELEX, an adaptive learning system designed to enhance procedural writing through personalized example-based learning. The core of our system is a multi-step example retrieval pipeline that selects a higher quality and contextually relevant example for each learner based on their unique input. We instantiate our system in the domain of cooking recipes. Specifically, we leverage a fine-tuned Large Language Model to predict the quality score of the learner's cooking recipe. Using this score, we retrieve recipes with higher quality from a vast database of over 180,000 recipes. Next, we apply BM25 to select the semantically most similar recipe in real-time. Finally, we use domain knowledge and regular expressions to enrich the selected example recipe with personalized instructional explanations. We evaluate RELEX in a 2 x 2 controlled study (personalized vs. non-personalized examples, reflective prompts vs. none) with 200 participants. Our results show that providing tailored examples contributes to better writing performance and user experience.

cs.HC↗

arXiv2026

PERK: Long-Context Reasoning as Test-Time Learning

Long-context reasoning requires accurately identifying relevant information in extensive, noisy input contexts. In this work, we propose PERK (Parameter Efficient Reasoning over Knowledge), a scalable approach for learning to encode long contexts using gradient updates at test time. Specifically, PERK employs two nested optimization loops in a meta-training phase. The inner loop rapidly encodes contexts into a low-rank adapter (LoRA) that serves as a parameter-efficient memory module for the base model. Concurrently, the outer loop learns to use the updated adapter to accurately recall and reason over relevant information from the encoded long context. Our evaluations on several long-context reasoning tasks show that PERK significantly outperforms the standard long-context finetuning, achieving average absolute performance gains of up to 20% for Qwen-2.5 (0.5B & 7B) on synthetic and real-world long-context reasoning. PERK also maintains its advantages across model scales and families. Compared to specialized long-context LLMs, PERK matches or surpasses their performance. Finally, our analyses show PERK is more robust to reasoning complexity, length extrapolation, and the positions of relevant information in contexts. https://perk-long-context.web.app

cs.CL↗

arXiv2026

Lingua Franca or Probing Artifact? Rethinking Latent Language in Multilingual LLMs

Latent language identification is often used to argue that multilingual language models route computation through language-specific states, such as English pivots. However, existing probes infer latent language from different signals, such as the geometry of hidden states or what can be decoded from intermediate representations. Since such claims shape conclusions about how models share and route information across languages, we ask whether these probes measure the same phenomenon or expose distinct aspects of multilingual computation. We study this question across model families, training regimes, domains, tasks, checkpoints, and up to 27 languages. We find that identification probes systematically disagree: the GMM-based representation probe, which draws evidence from hidden state geometry, shows earlier cross-lingual mixing, whereas decoding-based probes, which rely on output-space decodability, retain sharper language-specific and more English-biased signals. These differences track model multilinguality and training progression, but are comparatively stable across domains. Our results suggest a more cautious interpretation of latent language identification, where current probes expose different aspects of multilingual processing, rather than directly revealing a single internal lingua franca.

cs.CL↗

arXiv2026

CresOWLve: Benchmarking Creative Problem-Solving Over Real-World Knowledge

Creative problem-solving requires combining multiple cognitive abilities, including logical reasoning, lateral thinking, analogy-making, and commonsense knowledge, to discover insights that connect seemingly unrelated pieces of information. However, most existing benchmarks for large language models (LLMs) evaluate only specific components of this process. Moreover, many creativity-oriented benchmarks rely on artificially constructed brainteasers or contrived scenarios that do not reflect how creative problem-solving occurs in real-world settings. To address this gap, we introduce CresOWLve (containing 2,061 examples across five difficulty levels), a benchmark for evaluating creative problem-solving using puzzles grounded in real-world knowledge. Problems in CresOWLve require employing multiple creative thinking strategies, retrieving facts from diverse domains, and creatively combining them to arrive at a solution. Evaluating several frontier non-thinking and thinking LLMs, we show that CresOWLve remains highly challenging. Our analysis reveals a consistent performance gap: models perform substantially better on factual questions than on creative ones (up to -17% drop). While models can often retrieve the relevant knowledge, they struggle to form the non-obvious creative connections required to integrate the knowledge and arrive at the correct answer.

cs.CL↗

arXiv2026

Large Language Models Align with the Human Brain during Creative Thinking

Creative thinking is a fundamental aspect of human cognition, and divergent thinking-the capacity to generate novel and varied ideas-is widely regarded as its core generative engine. Large language models (LLMs) have recently demonstrated impressive performance on divergent thinking tests and prior work has shown that models with higher task performance tend to be more aligned to human brain activity. However, existing brain-LLM alignment studies have focused on passive, non-creative tasks. Here, we explore brain alignment during creative thinking using fMRI data from 170 participants performing the Alternate Uses Task (AUT). We extract representations from LLMs varying in size (270M-72B) and measure alignment to brain responses via Representational Similarity Analysis (RSA), targeting the creativity-related default mode and frontoparietal networks. We find that brain-LLM alignment is positively associated with model size (default mode network only) and with idea originality (both networks), with these relationships clearest early in the creative process. We further find that post-training objectives are associated with functionally selective differences in alignment: a creativity-optimized Llama-3.1-8B-Instruct retains alignment with high-creativity neural responses while lacking the positive low-creativity alignment present in other variants, but a reasoning-trained variant shows the opposite asymmetry, with negative alignment to high-creativity responses. Together, these results suggest that brain alignment offers an informative lens on how post-training relates to the neural geometry of human creative thought.

q-bio.NC↗

arXiv2026

DynaVieW: Schema-Guided World Modeling for Understanding Hierarchical Visual Dynamics

Multimodal LLMs struggle to systematically model the temporal evolution of visual scenes in videos or multi-image sequences. Such inputs require models to predict or simulate multiple levels of dynamic constituents, such as actions taken in the visual sequence, and the associated changes to the visual environment that result. To address this challenge, we propose a dynamic schema-guided world model, DynaVieW, optimized for visual dynamic prediction and simulation. DynaVieW achieves an in-depth understanding of visual dynamics by learning interleaved state-transition sequences, where states cover broad visual scenes from video keyframes, and transitions capture comprehensive dynamic constituents within a hierarchical schema. DynaVieW jointly models transition prediction and state simulation under a mixture-of-experts architecture, with a cross-expert selective attention and a schema token re-weighted loss, to ensure effective and robust learning. DynaVieW's understanding of visual dynamics boosts its downstream performance in visual narrative creation and world simulation, showing improved consistency, controllability, and instruction-following.

cs.LG↗

arXiv2026

Parity-Aware Byte-Pair Encoding: Improving Cross-lingual Fairness in Tokenization

Tokenization is the first -- and often least scrutinized -- step of most NLP pipelines. Standard algorithms for learning tokenizers rely on frequency-based objectives, which favor languages dominant in the training data and consequently leave lower-resource languages with tokenizations that are disproportionately longer, morphologically implausible, or even riddled with $ $ placeholders. This phenomenon ultimately amplifies computational and financial inequalities between users from different language backgrounds. To remedy this, we introduce Parity-aware Byte Pair Encoding (BPE), a variant of the widely-used BPE algorithm. At every merge step, Parity-aware BPE applies a fair-max rule that maximizes the compression gain of the currently worst-compressed language, trading a small amount of global compression for cross-lingual parity. We find empirically that Parity-aware BPE reduces tokenization inequality -- operationalized by the Gini coefficient of per-language token costs -- by up to 89% relative to Classical BPE. This comes with negligible impact on global compression rate and no evidence of systematic degradation in downstream LM performance.

cs.CL↗

arXiv2026

Diversity Matters: Revisiting Test-Time Compute in Vision-Language Models

Test-time compute (TTC) strategies have emerged as a lightweight approach to boost reasoning in large language models (LLMs). However, their application and benefits for vision-language models (VLMs) remain underexplored. We present a systematic study of TTC across seven VLMs and six benchmarks, specifically analyzing feature-based scoring and majority voting methods. We find that feature heuristics fail and voting yields only modest gains in single-model settings. We theoretically show that this limitation stems from a lack of prediction diversity: when outputs are highly correlated, voting provides little benefit. In contrast, multi-model ensembles offer richer diversity, yet standard majority voting fails to account for varying model capabilities. To address this, we propose Entropy-based TTC (ETTC), which selects the most confident prediction based on predictive entropy. Our method reduces to majority voting in the single-model case, but in model ensembles, it leverages confidence disparities to prioritize stronger models. We prove that ETTC outperforms majority voting under mild assumptions and empirically demonstrate that it consistently surpasses both voting and the best individual model. Crucially, our results show that smaller models can synergistically enhance larger ones, unlocking ensembling gains not achievable with standard strategies.

cs.LG↗

arXiv2026

Crosscoding Through Time: Tracking Emergence & Consolidation Of Linguistic Representations Throughout LLM Pretraining

Large language models (LLMs) learn non-trivial abstractions during pretraining, such as detecting irregular plural noun subjects. However, because traditional evaluation methods (e.g., benchmarking) fail to reveal how models acquire these concepts and capabilities, it is not well understood when and how these specific linguistic abilities emerge. To bridge this gap and better understand model training at the concept level, we use sparse crosscoders to discover and align features across model checkpoints. Using this approach, we track the evolution of linguistic features during pretraining. We train crosscoders between open-sourced checkpoint triplets with significant performance and representation shifts, and introduce a novel metric, Relative Indirect Effects (RelIE), to trace training stages at which individual features become causally important for task performance. We show that crosscoders can detect feature emergence, maintenance, and discontinuation during pretraining. Our approach is architecture-agnostic and scalable, offering a promising path toward more interpretable and fine-grained analysis of representation learning throughout pretraining.

cs.CL↗

arXiv2026

Do LLMs Game Formalization? Evaluating Faithfulness in Logical Reasoning

Formal verification guarantees proof validity but not formalization faithfulness. For natural-language logical reasoning, where models construct axiom systems from scratch without library constraints, this gap between valid proofs and faithful translations is especially acute. We investigate whether frontier models exploit this gap when generating Lean 4 proofs, a behavior we term formalization gaming. We evaluate GPT-5 and DeepSeek-R1 on 303 first-order logic problems (203 from FOLIO, 100 from Multi-LogiEval), comparing unified generation against a two-stage pipeline that separates formalization from proving. Despite compilation rates of 87-99%, we find no evidence of systematic gaming in unified generation: models prefer reporting failure over forcing proofs, even under prompting designed to encourage it. However, unfaithfulness that evades our detection signals may still occur. The two-stage pipeline reveals two distinct modes of unfaithfulness: GPT-5 fabricates axioms during proof generation, a reactive fallback detectable via cross-stage comparison, while DeepSeek-R1 mistranslates premises during formalization, producing internally consistent outputs that evade detection entirely. These findings show that high compilation rates or accuracies should not be equated with faithful reasoning. Code and data are available at https://github.com/koreankiwi99/formalization-gaming.

cs.AI↗