Activation-geometry method detects benchmark contamination in RL-trained models without training-data access

A preprint accepted at MAPR 2026 introduces "Excess Separability," a technique that reads residual-stream activation geometry inside transformer models to identify whether benchmark test items appeared in training data, requiring only inference-time access to intermediate activations and no training corpus disclosure. The method survives reinforcement-learning post-training, which prior work at ICLR 2026 showed can erase the perplexity-based signals that existing contamination detectors rely on. A concurrent method (LaRA, arXiv May 2026) pursues a similar internal-representation approach, suggesting a convergent research direction. For financial institutions evaluating AI vendor capability claims, the finding structurally weakens the assumption that published benchmark scores on static evaluations reflect genuine model capability, and creates an independent audit pathway for open-weight deployments that does not require lab cooperation. Also reported by: https://www.techtimes.com/articles/324511/20260814/benchmark-contamination-detection-inside-ai-models-new-method-survives-rl-post-training.htm