AI Paper Atlas

Research map / Large language models / Language, safety and interpretability

Beyond Benchmarks: Toward Causally Faithful Evaluation of Large Language Models

Zhengshuyuan Tian, Wanling Gao, Chuanxin Lan, Chenxi Wang, Lei Wang, Guoxin Kang, Zhengxin Yang, Yunyou Huang, Xuehai Hong, Jianfeng Zhan

ICML 2026 · Large language models › Language, safety and interpretability · approach: Pruning and circuits in LLMs

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
vary uncontrolled evaluation components (CoT, temperature, top p, presence penalty, max tokens) over 162 conditions (ablation)LLM evaluation · AIMEaccuracy: up to 70% variationMixed
evaluate under sampled SCS configurations instead of the default original settingDeepSeek · AIME'24accuracy: 0.367 -> 0.211Hurt
evaluate under sampled SCS configurations instead of the default original settingDoubao · AIME'24accuracy: 0.030 -> 0.254Improved
add augmented questions (A-SCS) to the SCS evaluationKimi · AIME'24accuracy: 0.082 -> 0.235Improved
add interaction of Shot and Multi Turn (ablation)GPT-3.5 · evaluation accuracysignificance: p = 0.58, p = 0.87 alone -> p = 0.01 interactionImproved
vary Question Format (ablation)DeepSeek · AIME'24 ANOVAvariance explained eta^2: 0.399643Improved
add chain-of-thought prompting (ablation)DeepSeek · AIME'24 ANOVAvariance explained eta^2: 0.080156Improved
vary max tokens (ablation)DeepSeek · AIME'24 ANOVAvariance explained eta^2: 0.028099Improved
vary temperature and top p jointly (ablation)DeepSeek · AIME'24 ANOVAvariance explained eta^2: 0.002082, p = 0.178No clear effect
vary Question Paraphrase and presence penalty (ablation)GPT-4.1 · AIME'24significance: p = 0.19 and 0.16No clear effect
compare models by interval estimates instead of default configurationpairwise model comparisons · AIME'24ranking agreement: 9 sign reversals, 7 missed differences, 3 illusory gaps (18 cases)Hurt
insert distractor sentences into a question (analogical transformation)nine LLMs · single seen questioncorrect solutions: 5 solved -> 0 solvedHurt

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy