Brittlebench: Quantifying LLM robustness via prompt sensitivity
FAIR at Meta(Meta FAIR);EPFL(瑞士联邦理工学院洛桑);Northeastern University(东北大学)
Angelika Romanou, Mark Ibrahim, Candace Ross, Kerem Oktar, Chantal Shaib, Samuel J. Bell, Anaelia Ovalle, Antoine Bosselut, Jesse Dodge, Koustuv Sinha, Adina Williams
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.
Quantifying Language Models' Sensitivity to Spurious Features in Prompt Design or: How I learned to start worrying about prompt formatting
Paul G. Allen School of Computer Science & Engineering, University of Washington(华盛顿大学保罗·G·艾伦计算机科学与工程学院);Allen Institute for Artificial Intelligence(艾伦人工智能研究所);University of California, Berkeley(加州大学伯克利分校)
Hedging and Non-Affirmation: Quantifying LLM Alignment on Questions of Human Rights
Google Deepmind(谷歌DeepMind);Massachusetts Institute of Technology(麻省理工学院);Independent Researcher(独立研究员);Google(谷歌);AI Accountability Lab, Trinity College Dublin(都柏林圣三一学院人工智能问责实验室)
Large Language Model Hacking: Quantifying the Hidden Risks of Using LLMs for Text Annotation
Bocconi University(博科尼大学);University of Zurich(苏黎世大学);University of Gothenburg(哥德堡大学);LIACS, Leiden University(莱顿大学高级计算机科学研究院);GESIS, Leibniz Institute for the Social Sciences(莱布尼茨社会科学研究所)
Quantifying Logical Consistency in Transformers via Query-Key Alignment
Skolkovo Institute of Science and Technology(斯科尔科沃科学技术研究所);Moscow Institute of Physics and Technology(莫斯科物理技术学院);CNRS(法国国家科学研究中心);Université Paris Cité(巴黎西岱大学);Artificial Intelligence Research Institute (AIRI)(人工智能研究所(AIRI))
Information Suppression in Large Language Models: Auditing, Quantifying, and Characterizing Censorship in DeepSeek
Thomas Lord Department of Computer Science, University of Southern California, USA(汤姆斯·劳德计算机科学系,南加州大学);Information Sciences Institute, University of Southern California, USA(信息科学研究所,南加州大学);Annenberg School of Communication, University of Southern California, USA(安纳伯格传播学院,南加州大学)
C-TPT: Calibrated Test-Time Prompt Tuning for Vision-Language Models via Text Feature Dispersion
Korea Advanced Institute of Science and Technology (KAIST)(韩国科学技术院(KAIST));University of Illinois at Urbana-Champaign (UIUC)(伊利诺伊大学厄巴纳-香槟分校(UIUC));Imperial College London(伦敦帝国学院)
Quantifying the Reconstructability of Astrophysical Methods with Large Language Models and Information Theory: A Case Study in Spectral Reconstruction
Department of Physics, University of Michigan, Ann Arbor, MI 48109, USA(密歇根大学物理系);Michigan Institute for Data and AI in Society, University of Michigan, Ann Arbor, MI 48109, USA(密歇根大学数据与人工智能在社会中的研究所);Independent Researcher, Taiwan(台湾独立研究员)
The Shape of Adversarial Influence: Characterizing LLM Latent Spaces with Persistent Homology
Microsoft Security Response Center(微软安全响应中心);AIDOS Lab, University of Fribourg(弗里堡大学AIDOS实验室);Department of Mathematics, Imperial College London(伦敦帝国理工学院数学系);Queen Mary University of London(伦敦大学量子玛丽学院);Imperial College London(伦敦帝国理工学院);Language Technology Lab, University of Cambridge(剑桥大学语言技术实验室);The Alan Turing Institute(艾伦·图灵研究所)
Sequential Bayesian Evaluation of Large Language Model Behavior
Department of Statistics, University of California, Irvine(统计系,加州大学伊文斯分校);Department of Computer Science, University of California, Irvine(计算机科学系,加州大学伊文斯分校)