The paper introduces Brittlebench, a framework and evaluation pipeline that measures how sensitive large language models (LLMs) are to prompt variations. By applying semantics‑preserving perturbations to popular benchmarks, the authors show that model performance can drop by up to 12% and that such perturbations can change model rankings in 63% of cases. The study finds that these input variations can explain up to half of the performance variance for a given model, underscoring the importance of robust evaluation methods.
By Angelika Romanou, Mark Ibrahim, Candace Ross, Kerem Oktar, Chantal Shaib, Samuel J. Bell, Anaelia Ovalle, Antoine Bosselut, Jesse Dodge, Koustuv Sinha, Adina Williams
The paper investigates how small lexical changes in prompts can cause large performance swings in large language models. Using a dataset of 132,000 prompt variants, the authors uncover a scaling law linking higher average task performance to lower variance and greater robustness. They identify domain-specific terminology and explicit action directives as key linguistic factors that stabilize prompts, and propose an automated Prompt-Refining Agent that reduces performance variance by 40.7% in code generation while maintaining or improving mean performance.
By Qipeng Xie, Zi Liang, Jiafei Wu, Yufei Chen, Weizheng Wang, Wenao Ma, Zhong Ming, Haiqin Yang, Kaishun Wu
MemToC is a controlled benchmark that tests how large language models resolve conflicts between their internal memory and tool outputs. It contains 6,504 episodes built from 542 factual questions, each paired with a model‑generated closed‑book answer and a tool return whose correctness is known, creating four distinct source‑correctness scenarios. Across five 7‑9B open‑weight models, tool responses overwhelmingly dominate closed‑book answers, and only a minority of instruction‑tuned models correctly retain a verified answer when the tool is wrong, while most follow a correct tool or repeat a wrong tool.
By Arseniy Varlamov, Rishat Zinnatullin, Elisei Rykov, Alexander Panchenko, Ilseyar Alimova
arXiv:2607.14109v2 Announce Type: replace
Abstract: Probing the capabilities of Large Language Models (LLMs) and building robust solutions for Multiple-Choice Question Answering (MCQA) remain central...
By Inder Preet, Shuxin Lin, Dhaval Patel
arXiv:2609.14157v1 Announce Type: new
Abstract: Large language models (LLMs) are increasingly deployed with external tools that extend what they can do beyond their own knowledge. Tools help on tasks...
By Saanvi Paturi, Arsen Kenzhebayev, Arham Sethi, Vyas Raina, Ivaxi Sheth, Vatsal Raina
arXiv:2610.19185v1 Announce Type: new
Abstract: Evaluation of large language models (LLMs) remains largely pre-scientific: it involves many design decisions, with little guidance on how to make them....
By Shenbin Qian, David Samuel, Jind\v{r}ich Helcl, Andrey Kutuzov, Stephan Oepen
PROOF is a benchmark that profiles the reliability of object-level facts in instruction-tuned language models by converting a frozen Wikidata snapshot into 18,486 English multiple-choice questions grounded in 11,779 semantic facts across 101 classes, 392 properties, and 14 domains. Each question includes an explicit "I don't know" option, a "No correct option" control, and nine controlled formulations, with 1,849 questions designed as no-correct-option traps. The study evaluates 18 open-weight model deployments on 166,374 prompts, revealing wide variability in factual accuracy, sensitivity to wording changes, and the impact of decoder perturbations.
By Andrei Chetvergov, Mikhail Solovev, Timofei Sivoraksha, Stepan Ukolov, Valeriia Kuschenko, Alexander Evseev, Sergey Bolovtsov
The paper introduces a new evaluation framework for confidence estimation in large language models, focusing on three properties: robustness to prompt changes, stability across semantically equivalent answers, and sensitivity to semantically different answers. It demonstrates that existing confidence estimation methods perform well on robustness and stability but often fail to detect differences in answer meaning, revealing gaps in current evaluation practices. The framework aims to guide the selection of confidence estimators for practical applications.
By Yuxi Xia, Dennis Ulmer, Terra Blevins, Yihong Liu, Hinrich Sch\"utze, Benjamin Roth
The paper introduces the "Prompt Lottery" phenomenon, where paraphrastic rephrasings of judge prompts in LLM-as-a-Judge (LLMaJ) lead to inconsistent judgments and can even reverse model rankings. A large-scale study across various tasks, judge families, and model backbones demonstrates that these prompt variations significantly alter evaluation outcomes. A user study shows that LLMaJ judges are 6–9 times more sensitive to prompt rephrasings than humans, and that aggregating multiple prompts improves robustness by focusing on consistent judgments.
By Ilija Subasic, Brett Levert, Andrew Rabinovich, Sajjad Pakdamansavoji, Zhao Chen
The paper evaluates how robust large language models (LLMs) are when using retrieval‑augmented generation (RAG) in practical settings. It investigates whether RAG always outperforms non‑RAG approaches, whether adding more retrieved documents helps, and whether the order of documents matters, using a benchmark of 1,891 samples across five datasets and three task categories. Experiments with 11 LLMs show generally high retrieval robustness, but performance varies by task and prompting strategy, indicating that adopting RAG should be considered on a case‑by‑case basis.
By Shuyang Cao, Karthik Radhakrishnan, David Rosenberg, Steven Lu, Pengxiang Cheng, Lu Wang, Shiyue Zhang
arXiv:2604. 09497v2 Announce Type: replace-cross Abstract: Accurate evaluation is central to the large language model (LLM) ecosystem, guiding model selection and downstream adoption across diverse use cases.
By Hippolyte Gisserot-Boukhlef, Nicolas Boizard, Emmanuel Malherbe, C\'eline Hudelot, Pierre Colombo
The paper introduces soft‑prompt tuning as a lightweight, automated method for evaluating large language models (LLMs). By optimizing only ten soft‑prompt vectors—about 0.0006% of a 7B model’s parameters—within 150 steps, the approach adapts models to specific benchmark formats, improving format compliance while preserving knowledge accuracy. Experiments across seven models and eleven datasets show that this method matches strong evaluation‑harness baselines, enhances evaluation signals on challenging datasets, and yields cleaner outputs that align better with exact‑match scoring and independent LLM judges.
By Selen Erkan, Bastian Boll, Kristian Kersting, Bj\"orn Deiseroth, Letitia Parcalabescu