Phare: A Safety Probe for Large Language Models

Pierre Le Jeune; Benoît Malézieux; Weixuan Xiao; Matteo Dora

Home/Archive/arXiv

arXiv · 2505.11365

Phare: A Safety Probe for Large Language Models

Pierre Le Jeune·Benoît Malézieux·Weixuan Xiao·Matteo Dora

View original source ↗Read source document ↗Export metadata ↓

Abstract

Ensuring the safety of large language models (LLMs) is critical for responsible deployment, yet existing evaluations often prioritize performance over identifying failure modes. We introduce Phare, a multilingual diagnostic framework to probe and evaluate LLM behavior across three critical dimensions: hallucination and reliability, social biases, and harmful content generation. Our evaluation of 17 state-of-the-art LLMs reveals patterns of systematic vulnerabilities across all safety dimensions, including sycophancy, prompt sensitivity, and stereotype reproduction. By highlighting these specific failure modes rather than simply ranking models, Phare provides researchers and practitioners with actionable insights to build more robust, aligned, and trustworthy language systems.

Explore related subjects

cs.CY cs.AI cs.CL cs.CR

Keep this discovery

Explore connections, maps & timelines

BibTeX RIS

Pierre Le Jeune, Benoît Malézieux, Weixuan Xiao, Matteo Dora. 2025-05-26. Phare: A Safety Probe for Large Language Models. https://arxiv.org/abs/2505.11365

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

Frontier Lag: A Bibliometric Audit of Capability Misrepresentation in Academic AI Evaluation

LLM evaluations in applied domains tend to reflect models that were already outclassed at time of publication. We observe a publication elicitation gap: the distance between the AI systems generating the results reported in an academic paper and the AI systems that a current reader of that paper would reasonably assume are being referenced. We systematically sweep OpenAlex from 2022-01-01 to 2026-04-01 (n = 112,303 LLM keyword matches). Then, we identify what models were evaluated (n = 18,574 admissible records). We then rank each evaluated LLM against a frontier LLM based on the Epoch AI Capabilities Index (ECI), an aggregate LLM capability score. At time of evaluation, the median paper is evaluating models that are behind frontier LLMs in capability, with a median gap of +10.85 ECI (H1; n = 12,312). This gap is growing, increasing at a rate of +5.53 ECI per year (H2, nominal 95% CI [+5.03, +5.83]). The sign holds even in the absence of any imputation for evaluation date. In papers (n = 728) where the date of evaluation is explicit and the model in question can be resolved to an ECI score, the median gap for H1 is +5.01 ECI. An explicitly stated evaluation date can be found in only 18.4% of full-text papers. After correction, in 52.5% (95% CI: [48.2, 56.9]) of abstracts in our audit, conclusions are stated at the class level ("AI") rather than the model level. For papers about reasoning models, only 3.2% of abstracts and 21.2% of full-text articles disclose the reasoning mode status of the models used (H4). We propose a solution to this problem that is distributed among authors, editors, and funders. First, reporting from authors. VERSIO-AI v1.2 is a proposed 13-item checklist to cover the configuration surface described herein. Second, enforcement from journal editors and peer reviewers. Third, conditioning grants on disclosure and providing API access.

cs.CY↗

arXiv2026

The Cross-Section of Stock Returns and AI Exposure

We study 380 trillion tokens of realized AI consumption across more than four hundred LLMs. We build a high-frequency AI factor and show that a long-short strategy based on firms' AI exposure earns significantly positive returns. The average strategy return is larger based on intensive, frontier-oriented AI consumption but smaller based on casual or open-weight usage. Internationally, the return spread is significant in developed countries but insignificant in emerging markets. Examining occupational AI exposure, we find more positive exposure in occupations intensive in nonroutine interactive tasks and more negative exposure in those intensive in nonroutine analytical tasks.

cs.CY↗

arXiv2026

The queer Hero versus the Fool bias of the queer trait: An archetypometric analysis of the collective portrayal of queerness in fictional stories

Visibility in media is pivotal for identity development and for broadening societal views of gender and sexuality. Queer representation has increased in recent years, yet damaging stereotypes and tropes persist. Here, we focus on queer portrayal and its perception by audiences in fictional stories (television, film, and literature) by studying characters by their quantified archetypes which are operationalizations of common conceptions such as Hero, Diva, and Outcast. We use the archetypometrics and Fandom's LGBTQIA+ datasets to study samples of fictional characters along the trait differential spanning straight to queer. We find, quantify, and explain a seeming paradox. The characters with the highest queer score present positive primary archetypes and are typically Heroes rather than Fools, Angels rather than Demons, and Adventurers rather than Traditionalists. But evaluation across many stories for the straight-queer trait itself reveals a strong collective-writing bias towards Fool (away from Hero) and no meaningful loading for the other two dimensions. Our analysis offers a population-scale view of the complexities of queer portrayal, while also pointing to risks in blindly training on many-authored story corpora.

cs.CY↗