Reproduce “On The Fragility of Benchmark Contamination Detection in Reasoning Models” on OpenResearch

About this paper

Han Wang, Haoyu Li, Brian Ko, Huan Zhang

Research reveals that existing benchmark contamination detection methods are ineffective for Large Reasoning Models (LRMs), as Reinforcement Learning (RL) training can conceal contamination introduced during model evolution. Furthermore, extensive Chain-of-Thought (CoT) contamination on advanced LRMs substantially boosts performance while remaining largely undetectable, challenging detection methods reliant on memorization.

Key insights

• Reinforcement Learning (RL) training, particularly using PPO-style objectives like GRPO, can conceal benchmark contamination introduced during Supervised Fine-Tuning (SFT) without the model "forgetting" the data.
• The concealment mechanism is attributed to the importance sampling and clipping terms within PPO-style RL objectives, which reduce the negative log-likelihood gap between member and non-member samples.
• Advanced Large Reasoning Models (LRMs) generalize from Chain-of-Thought (CoT) contaminated data, increasing confidence on unseen but similar samples, which undermines the core memorization assumption of current detection methods.

Summary by alphaXiv.