arXiv subjects

Antoine Bosselut

Publications and source records attributed to Antoine Bosselut.

At least 37 records · Page 2Linked to original sources

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

ConLID: Supervised Contrastive Learning for Low-Resource Language Identification

Language identification (LID) is a critical step in curating multilingual LLM pretraining corpora from web crawls. While many studies on LID model training focus on collecting diverse training data to improve performance, low-resource languages -- often limited to single-domain data, such as the Bible -- continue to perform poorly. To resolve these imbalance and bias issues, we propose a novel supervised contrastive learning (SCL) approach to learn domain-invariant representations for low-resource languages. We show that our approach improves LID performance on out-of-domain data for low-resource languages by 3.2 percentage points, while maintaining its performance for the high-resource languages.

cs.CL↗

arXiv2026

AI Meets Mathematics Education: A Case Study on Supporting an Instructor in a Large Mathematics Class with Context-Aware AI

Large-enrollment university courses face persistent challenges in providing timely and scalable instructional support. While generative AI holds promise, its effective use depends on reliability and pedagogical alignment. We present a human-centered case study of AI-assisted support in a Calculus I course, implemented in close collaboration with the course instructor. We developed a system to answer students' questions on a discussion forum, fine-tuning a lightweight language model on 2,588 historical student-instructor interactions. The model achieved 75.3% accuracy on a benchmark of 150 representative questions annotated by five instructors, and in 36% of cases, its responses were rated equal to or better than instructor answers. Post-deployment student survey (N = 105) indicated that students valued the alignment of the responses with the course materials and their immediate availability, while still relying on the instructor verification for trust. We highlight the importance of hybrid human-AI workflows for safe and effective course support.

cs.CY↗

arXiv2026

AbstRaL: Augmenting LLMs' Reasoning by Reinforcing Abstract Thinking

Recent studies have shown that large language models (LLMs), especially smaller ones, often lack robustness in grade school math (GSM) reasoning. In particular, they tend to experience performance drops when faced with distribution shifts, such as changes to numerical or nominal variables, or insertions of distracting clauses. A possible strategy to address this involves generating synthetic data to further "instantiate" reasoning problems on potential variations. In this work, we instead focus on the strategy of "abstracting" reasoning problems. This not only helps counteract distribution shifts but also facilitates the connection to symbolic tools for deriving solutions. Focusing on GSM, we find that this abstraction process is better acquired through reinforcement learning (RL) than just supervised fine-tuning, which often fails to produce faithful abstractions. Our method, AbstRaL -- which promotes abstract reasoning in LLMs using RL on granular abstraction data -- significantly mitigates performance degradation on recent GSM perturbation benchmarks. Besides, improving GSM robustness via AbstRaL is shown to also implicitly benefit LLMs' capabilities on OOD mathematical and general reasoning tasks, indicating that abstract thinking broadly enables better generalizability.

cs.CL↗

arXiv2026

Buy versus Build an LLM: A Decision Framework for Governments

Large Language Models (LLMs) represent a new frontier of digital infrastructure that can support a wide range of public-sector applications, from general purpose citizen services to specialized and sensitive state functions. When expanding AI access, governments face a set of strategic choices over whether to buy existing services, build domestic capabilities, or adopt hybrid approaches across different domains and use cases. These are critical decisions especially when leading model providers are often foreign corporations, and LLM outputs are increasingly treated as trusted inputs to public decision-making and public discourse. In practice, these decisions are not intended to mandate a single approach across all domains; instead, national AI strategies are typically pluralistic, with sovereign, commercial and open-source models coexisting to serve different purposes. Governments may rely on commercial models for non-sensitive or commodity tasks, while pursuing greater control for critical, high-risk or strategically important applications. This paper provides a strategic framework for making this decision by evaluating these options across dimensions including sovereignty, safety, cost, resource capability, cultural fit, and sustainability. Importantly, "building" does not imply that governments must act alone: domestic capabilities may be developed through public research institutions, universities, state-owned enterprises, joint ventures, or broader national ecosystems. By detailing the technical requirements and practical challenges of each pathway, this work aims to serve as a reference for policy-makers to determine whether a buy or build approach best aligns with their specific national needs and societal goals.

cs.CY↗

arXiv2026

Evaluating Language Model Agency through Negotiations

We introduce an approach to evaluate language model (LM) agency using negotiation games. This approach better reflects real-world use cases and addresses some of the shortcomings of alternative LM benchmarks. Negotiation games enable us to study multi-turn, and cross-model interactions, modulate complexity, and side-step accidental evaluation data leakage. We use our approach to test six widely used and publicly accessible LMs, evaluating performance and alignment in both self-play and cross-play settings. Noteworthy findings include: (i) only closed-source models tested here were able to complete these tasks; (ii) cooperative bargaining games proved to be most challenging to the models; and (iii) even the most powerful models sometimes "lose" to weaker opponents

cs.CL↗

arXiv2026

Tracking the Limits of Knowledge Propagation: How LLMs Fail at Multi-Step Reasoning with Conflicting Knowledge

A common solution for mitigating outdated or incorrect information in Large Language Models (LLMs) is to provide updated facts in-context or through knowledge editing. However, these methods introduce knowledge conflicts when the knowledge update fails to overwrite the model's parametric knowledge, which propagate to faulty reasoning. Current benchmarks for this problem, however, largely focus only on single knowledge updates and fact recall without evaluating how these updates affect downstream reasoning. In this work, we introduce TRACK (Testing Reasoning Amid Conflicting Knowledge), a new benchmark for studying how LLMs propagate new knowledge through multi-step reasoning when it conflicts with the model's initial parametric knowledge. Spanning three reasoning-intensive scenarios (WIKI, CODE, and MATH), TRACK introduces multiple, realistic conflicts to mirror real-world complexity. Our results on TRACK reveal that providing updated facts to models for reasoning can worsen performance compared to providing no updated facts to a model, and that this performance degradation exacerbates as more updated facts are provided. We show this failure stems from both inability to faithfully integrate updated facts, but also flawed reasoning even when knowledge is integrated. TRACK provides a rigorous new benchmark to measure and guide future progress on propagating conflicting knowledge in multi-step reasoning.

cs.AI↗

arXiv2025

Measuring what Matters: Construct Validity in Large Language Model Benchmarks

Evaluating large language models (LLMs) is crucial for both assessing their capabilities and identifying safety or robustness issues prior to deployment. Reliably measuring abstract and complex phenomena such as 'safety' and 'robustness' requires strong construct validity, that is, having measures that represent what matters to the phenomenon. With a team of 29 expert reviewers, we conduct a systematic review of 445 LLM benchmarks from leading conferences in natural language processing and machine learning. Across the reviewed articles, we find patterns related to the measured phenomena, tasks, and scoring metrics which undermine the validity of the resulting claims. To address these shortcomings, we provide eight key recommendations and detailed actionable guidance to researchers and practitioners in developing LLM benchmarks.

cs.CL↗

arXiv2025

Reliable Evaluation and Benchmarks for Statement Autoformalization

Evaluating statement autoformalization, translating natural language mathematics into formal languages like Lean 4, remains a significant challenge, with few metrics, datasets, and standards to robustly measure progress. In this work, we present a comprehensive approach combining improved metrics, robust benchmarks, and systematic evaluation, to fill this gap. First, we introduce BEq+, an automated metric that correlates strongly with human judgment, along with ProofNetVerif, a new dataset for assessing the quality of evaluation metrics, containing 3,752 annotated examples. Second, we develop two new autoformalization benchmarks: ProofNet#, a corrected version of ProofNet, and RLM25, with 619 new pairs of research-level mathematics from six formalization projects. Through systematic experimentation across these benchmarks, we find that current techniques can achieve up to 45.1% accuracy on undergraduate mathematics but struggle with research-level content without proper context. Our work establishes a reliable foundation for evaluating and advancing autoformalization systems.

cs.CL↗

arXiv2025

RLMEval: Evaluating Research-Level Neural Theorem Proving

Despite impressive results on curated benchmarks, the practical impact of large language models (LLMs) on research-level neural theorem proving and proof autoformalization is still limited. We introduce RLMEval, an evaluation suite for these tasks, focusing on research-level mathematics from real-world Lean formalization projects. RLMEval targets the evaluation of neural theorem proving and proof autoformalization on challenging research-level theorems by leveraging real Lean Blueprint formalization projects. Our evaluation of state-of-the-art models on RLMEval, comprising 613 theorems from 6 Lean projects, reveals a significant gap: progress on existing benchmarks does not readily translate to these more realistic settings, with the best model achieving only a 10.3 % pass rate. RLMEval provides a new, challenging benchmark designed to guide and accelerate progress in automated reasoning for formal mathematics.

cs.CL↗

arXiv2025

Revisiting Multilingual Data Mixtures in Language Model Pretraining

The impact of different multilingual data mixtures in pretraining large language models (LLMs) has been a topic of ongoing debate, often raising concerns about potential trade-offs between language coverage and model performance (i.e., the curse of multilinguality). In this work, we investigate these assumptions by training 1.1B and 3B parameter LLMs on diverse multilingual corpora, varying the number of languages from 25 to 400. Our study challenges common beliefs surrounding multilingual training. First, we find that combining English and multilingual data does not necessarily degrade the in-language performance of either group, provided that languages have a sufficient number of tokens included in the pretraining corpus. Second, we observe that using English as a pivot language (i.e., a high-resource language that serves as a catalyst for multilingual generalization) yields benefits across language families, and contrary to expectations, selecting a pivot language from within a specific family does not consistently improve performance for languages within that family. Lastly, we do not observe a significant "curse of multilinguality" as the number of training languages increases in models at this scale. Our findings suggest that multilingual data, when balanced appropriately, can enhance language model capabilities without compromising performance, even in low-resource settings

cs.CL↗

arXiv2025

CAVE: Detecting and Explaining Commonsense Anomalies in Visual Environments

Humans can naturally identify, reason about, and explain anomalies in their environment. In computer vision, this long-standing challenge remains limited to industrial defects or unrealistic, synthetically generated anomalies, failing to capture the richness and unpredictability of real-world anomalies. In this work, we introduce CAVE, the first benchmark of real-world visual anomalies. CAVE supports three open-ended tasks: anomaly description, explanation, and justification; with fine-grained annotations for visual grounding and categorizing anomalies based on their visual manifestations, their complexity, severity, and commonness. These annotations draw inspiration from cognitive science research on how humans identify and resolve anomalies, providing a comprehensive framework for evaluating Vision-Language Models (VLMs) in detecting and understanding anomalies. We show that state-of-the-art VLMs struggle with visual anomaly perception and commonsense reasoning, even with advanced prompting strategies. By offering a realistic and cognitively grounded benchmark, CAVE serves as a valuable resource for advancing research in anomaly detection and commonsense reasoning in VLMs.

cs.CV↗

arXiv2025

GaLLoP: Gradient-based Sparse Learning on Low-Magnitude Parameters

Sparse fine-tuning techniques adapt LLMs to downstream tasks by only tuning a sparse subset of model parameters. However, the effectiveness of sparse adaptation depends on optimally selecting the model parameters to be fine-tuned. In this work, we introduce a novel sparse fine-tuning technique named GaLLoP: Gradient-based Sparse Learning on Low-Magnitude Parameters, which fine-tunes only those model parameters which have the largest gradient magnitudes on downstream tasks and the smallest pre-trained magnitudes, intuitively prioritizing parameters that are highly task-relevant, but minimally disruptive to pre-trained knowledge. Our experimentation with LLaMA3 8B and Gemma 2B as base models shows that GaLLoP consistently improves or matches the in-distribution as well as out-of-distribution performance obtained via the usage of other leading parameter-efficient fine-tuning techniques, including LoRA, DoRA, and SAFT. Our analysis demonstrates that GaLLoP mitigates catastrophic forgetting and memorization of task data, as important pre-trained parameters remain unchanged, and stabilizes performance relative to other fine-tuning techniques, robustly generalizing across most random seeds.

cs.LG↗

arXiv2025

LLMs Are In-Context Bandit Reinforcement Learners

Large Language Models (LLMs) excel at in-context learning (ICL), a supervised learning technique that relies on adding annotated examples to the model context. We investigate a contextual bandit version of in-context reinforcement learning (ICRL), where models learn in-context, online, from external reward, instead of supervised data. We show that LLMs effectively demonstrate such learning, and provide a detailed study of the phenomena, experimenting with challenging classification tasks and models of sizes from 500M to 70B parameters. This includes identifying and addressing the instability of the process, demonstrating learning with both semantic and abstract labels, and showing scaling trends. Our findings highlight ICRL capabilities in LLMs, while also underscoring fundamental limitations in their implicit reasoning about errors.

cs.CL↗

arXiv2025

From Language to Cognition: How LLMs Outgrow the Human Language Network

Large language models (LLMs) exhibit remarkable similarity to neural activity in the human language network. However, the key properties of language shaping brain-like representations, and their evolution during training as a function of different tasks remain unclear. We here benchmark 34 training checkpoints spanning 300B tokens across 8 different model sizes to analyze how brain alignment relates to linguistic competence. Specifically, we find that brain alignment tracks the development of formal linguistic competence -- i.e., knowledge of linguistic rules -- more closely than functional linguistic competence. While functional competence, which involves world knowledge and reasoning, continues to develop throughout training, its relationship with brain alignment is weaker, suggesting that the human language network primarily encodes formal linguistic structure rather than broader cognitive functions. We further show that model size is not a reliable predictor of brain alignment when controlling for feature size and find that the correlation between next-word prediction, behavioral alignment and brain alignment fades once models surpass human language proficiency. Finally, using the largest set of rigorous neural language benchmarks to date, we show that language brain alignment benchmarks remain unsaturated, highlighting opportunities for improving future models. Taken together, our findings suggest that the human language network is best modeled by formal, rather than functional, aspects of language.

cs.CL↗

arXiv2025

Creative Preference Optimization

While Large Language Models (LLMs) have demonstrated impressive performance across natural language generation tasks, their ability to generate truly creative content-characterized by novelty, diversity, surprise, and quality-remains limited. Existing methods for enhancing LLM creativity often focus narrowly on diversity or specific tasks, failing to address creativity's multifaceted nature in a generalizable way. In this work, we propose Creative Preference Optimization (CrPO), a novel alignment method that injects signals from multiple creativity dimensions into the preference optimization objective in a modular fashion. We train and evaluate creativity-augmented versions of several models using CrPO and MuCE, a new large-scale human preference dataset spanning over 200,000 human-generated responses and ratings from more than 30 psychological creativity assessments. Our models outperform strong baselines, including GPT-4o, on both automated and human evaluations, producing more novel, diverse, and surprising generations while maintaining high output quality. Additional evaluations on NoveltyBench further confirm the generalizability of our approach. Together, our results demonstrate that directly optimizing for creativity within preference frameworks is a promising direction for advancing the creative capabilities of LLMs without compromising output quality.

cs.CL↗

arXiv2025

Can Performant LLMs Be Ethical? Quantifying the Impact of Web Crawling Opt-Outs

The increasing adoption of web crawling opt-outs by copyright holders of online content raises critical questions about the impact of data compliance on large language model (LLM) performance. However, little is known about how these restrictions (and the resultant filtering of pretraining datasets) affect the capabilities of models trained using these corpora. In this work, we conceptualize this effect as the $\textit{data compliance gap}$ (DCG), which quantifies the performance difference between models trained on datasets that comply with web crawling opt-outs, and those that do not. We measure the data compliance gap in two settings: pretraining models from scratch and continual pretraining from existing compliant models (simulating a setting where copyrighted data could be integrated later in pretraining). Our experiments with 1.5B models show that, as of January 2025, compliance with web data opt-outs does not degrade general knowledge acquisition (close to 0\% DCG). However, in specialized domains such as biomedical research, excluding major publishers leads to performance declines. These findings suggest that while general-purpose LLMs can be trained to perform equally well using fully open data, performance in specialized domains may benefit from access to high-quality copyrighted sources later in training. Our study provides empirical insights into the long-debated trade-off between data compliance and downstream model performance, informing future discussions on AI training practices and policy decisions. Our website is available at https://data-compliance.github.io/.

cs.CL↗

arXiv2025

GeoExplorer: Active Geo-localization with Curiosity-Driven Exploration

Active Geo-localization (AGL) is the task of localizing a goal, represented in various modalities (e.g., aerial images, ground-level images, or text), within a predefined search area. Current methods approach AGL as a goal-reaching reinforcement learning (RL) problem with a distance-based reward. They localize the goal by implicitly learning to minimize the relative distance from it. However, when distance estimation becomes challenging or when encountering unseen targets and environments, the agent exhibits reduced robustness and generalization ability due to the less reliable exploration strategy learned during training. In this paper, we propose GeoExplorer, an AGL agent that incorporates curiosity-driven exploration through intrinsic rewards. Unlike distance-based rewards, our curiosity-driven reward is goal-agnostic, enabling robust, diverse, and contextually relevant exploration based on effective environment modeling. These capabilities have been proven through extensive experiments across four AGL benchmarks, demonstrating the effectiveness and generalization ability of GeoExplorer in diverse settings, particularly in localizing unfamiliar targets and environments.

cs.CV↗