LaRA: Layer-wise Representation Analysis for Detecting Data Contamination in RL Post-Training

Reinforcement learning (RL) improves reasoning in large language models (LLMs) but can also induce contamination through reward-driven memorization. Existing contamination detection methods mainly rely on output-level signals, such as likelihood or entropy, which become unstable after RL post-training. We propose LaRA, a layer-wise representation analysis framework for contamination detection in RL-trained LLMs. LaRA introduces three complementary metrics—Representation Shift Magnitude (RSM), Directional Collapse (DC), and Representation Stability Index (RSI)—to measure perturbation sensitivity, directional concentration, and local representation variability under controlled perturbations. Our analyses show that contamination induces progressive geometric deviations across layers, characterized by amplified perturbation sensitivity, abnormal directional concentration dynamics, and unstable local variability. Based on these observations, we develop a layer-aware detection protocol that aggregates representation-level deviations across layers and metrics. Experiments on RL-trained reasoning models show that LaRA consistently improves contamination detection over existing output-level baselines.