AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

On The Fragility of Benchmark Contamination Detection in Reasoning Models

Han Wang, Haoyu Li, Brian Ko, Huan Zhang

ICLR 2026 · Reinforcement learning › RL from feedback · approach: Scaling RL with verifiable rewards · 8 citations (OpenAlex)

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
run GRPO with clean data after SFT contamination (Min-K%) (ablation)SFT-contaminated Qwen2.5-7B · 6 reasoning benchmarksAUROC (%): 74.96 -> 61.27Hurt
run GRPO with clean data (LiRA) (ablation)SFT-contaminated Qwen2.5-7B · 6 reasoning benchmarksAUROC (%): 89.13 -> 80.14Hurt
run GRPO with clean data (Max-K%) (ablation)SFT-contaminated Qwen2.5-7B · 6 reasoning benchmarksAUROC (%): 69.83 -> 52.35Hurt
GRPO on clean & member data instead of clean only (ablation)SFT-contaminated Qwen2.5-7B · 6 reasoning benchmarksAUROC (%): 61.27 -> 58.54Hurt
add clipping to RAFT++ objective (ablation)RAFT++ · 6 reasoning benchmarksLoss AUROC (%): 74.39 -> 57.58Hurt
add clipping to GRPO objective (ablation)GRPO · 6 reasoning benchmarksLoss AUROC (%): 73.28 -> 61.26Hurt
plain rejection sampling (RAFT) after SFT contamination (ablation)SFT-contaminated Qwen2.5-7B · 6 reasoning benchmarksLoss AUROC (%): 75.48 -> 77.51No clear effect
SFT contamination on Qwen2.5-7B-Instruct before RLclean SFT · 6 reasoning benchmarkspass@1 avg (%): 38.41 -> 47.23Improved
RL contamination (benchmark questions with rewards) vs clean RLQwen2.5-7B SFT · 6 reasoning benchmarkspass@1 avg (%): 45.55 -> 44.96No clear effect
extensive SFT contamination with CoT on DS Llama-8BDeepSeek-R1-Distill-Llama-8B · 6 reasoning benchmarkspass@1 avg (%): 48.38 -> 60.14Improved
extensive SFT contamination with CoT on DS Qwen-7BDeepSeek-R1-Distill-Qwen-7B · 6 reasoning benchmarkspass@1 avg (%): 54.42 -> 57.10Improved
detect final-stage CoT contamination (Max-K%)contaminated DS Qwen-14B · 6 reasoning benchmarksAUROC (%): 51.32No 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