AI Paper Atlas

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

CapBencher: Give Your LLM Benchmark a Built-in Alarm for Test-Set Overfitting

Takashi Ishida, Thanawat Lodkaew, Ikko Yamane

ICML 2026 · Large language models › Language, safety and interpretability · approach: Uncertainty quantification for 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
cap benchmark accuracy at 50% Bayes accuracyCapBencher · multiple LLM benchmarksranking agreement (Kendall's tau): Kendall's tau 0.92No clear effect
cap small benchmark (GPQA diamond, n=198)CapBencher · GPQA diamondKendall's tau: 0.93 (MMLU-Pro) -> 0.67 (GPQA diamond)Hurt
use larger model for contamination detectionCapBencher binomial test · continued pretraining on capped benchmarksepochs to detect contaminationImproved
lower Bayes accuracy (50% -> 25% -> 10%) (ablation)CapBencher on Llama 3.2-3B-Instruct · GSM8Kepochs to detect contamination: epoch 6 (50%), 5 (25%), 4 (10%)Improved
lower Bayes accuracy (larger number of answer options) (ablation)CapBencher · accuracy estimationroom to track LLM improvementHurt
require reasoning in response before randomized answer (ablation)CapBencher multiple-choice · uncontaminated model, multiple-choiceadherence to randomization instructionImproved
cross-lingual contamination detection with reasoning in responseCapBencher on Qwen 2.5-3B-Instruct · GSM8K translated, with reasoningcontamination detectionHurt
compare CapBencher to Min-k% and Min-k%++ detectioncontamination detection · GPQA and othersdetection successImproved
evolutionary model merging optimized on capped GSM8KCapBencher binomial test · capped GSM8K (50%)accuracy (%): 41.02 -> 56.52Improved
reverse-engineer true answers from randomized answer with P1/P2 promptscapped benchmarks, Qwen 2.5-14B-Instruct and GPT-5 models · nine benchmarksrecovered accuracyNo clear effect
reverse engineering with P3 on HumanEvalCapBencher disclosure strategy · HumanEvalaccuracy (%): 82.9 -> 82.3No clear effect

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