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

Reinforcement learning (RL) post-training has shown to improve reasoning in large language models (LLMs). However, there has been little exploration on the problem of data contamination in RL post-training, potentially undermining generalization and evaluation reliability of the training process itself. Existing detection methods primarily rely on output-level signals such as likelihood or entropy, which become unreliable for RL-trained models since RL shapes behavior through trajectory-level rewards rather than token likelihoods. We propose LaRA, a layer-wise representation analysis framework for detecting contamination in RL post-trained LLMs. LaRA introduces three complementary metrics, measuring perturbation sensitivity, directional collapse, and local representation rigidity under controlled perturbations. We find that contamination produces progressive geometric deviations across layers, including amplified perturbation sensitivity, stronger directional collapse, and enhanced local rigidity. Based on our findings, we also develop a contamination detection protocol that aggregates representation-level deviations across layers and metrics. Experiments on RL-trained reasoning models show that our protocol outperforms existing output-level baselines for contamination detection.

📋 한 줄 요약

**[RL Post-Training / Data Contamination 탐지]** LaRA가 layer-wise representation 분석으로 RL post-trained LLM의 contamination 탐지 — perturbation sensitivity·directional collapse·local rigidity 3 지표로 layer 간 누진 geometric deviation 측정, output-level baseline 능가.

🎯 핵심 기여도

RL post-training이 LLM 추론 개선에 효과적이지만 RL post-training의 data contamination 문제는 거의 탐구 안 됨 — 학습 과정 자체의 일반화·평가 신뢰성을 잠재적 저해.
기존 탐지 방법이 주로 likelihood·entropy 같은 output-level 신호 의존, RL이 token likelihood가 아닌 trajectory-level reward로 행동 형성하므로 RL-trained 모델에 unreliable.
LaRA — RL post-trained LLM의 contamination 탐지를 위한 layer-wise representation analysis 프레임워크.
3 보완 지표 — perturbation sensitivity·directional collapse·local representation rigidity — 통제된 perturbation 하에서 측정. Contamination이 layer 전반에 progressive geometric deviation 유발(amplified perturbation sensitivity·stronger directional collapse·enhanced local rigidity) 발견. 이 발견 기반 layer·지표 전반의 representation deviation 집계하는 탐지 프로토콜 개발.

💡 핵심 아이디어

RL post-training의 contamination은 token-level output 신호에 잘 드러나지 않고 hidden representation의 geometric deviation(perturbation sensitivity·directional collapse·local rigidity)으로 progressive하게 나타나며, layer-wise 분석이 trajectory-reward 형성 행동에 적합한 탐지 신호다.

🔬 기술적 접근법

**방법론**: LaRA — layer-wise representation analysis.
**핵심 기법**: (1) 통제된 perturbation 하 hidden representation 분석, (2) Perturbation sensitivity 지표(작은 입력 변화에 대한 표현 변화 amplification), (3) Directional collapse(표현 방향 다양성 감소), (4) Local representation rigidity(국소 표현의 유연성 감소), (5) Layer·지표 전반 deviation 집계 탐지 프로토콜.

📊 주요 결과

Contamination이 layer 전반에 progressive geometric deviation 유발 확인.
Amplified perturbation sensitivity·stronger directional collapse·enhanced local rigidity 3 시그니처 일관 관측.
RL-trained reasoning model에서 LaRA가 기존 output-level baseline 능가.

💭 의의 및 한계

**의의**: RL post-training의 무시되던 contamination 문제 정식화, output-level 신호의 trajectory-reward 시나리오 한계 진단, layer-wise representation 분석이라는 일반화 가능 패러다임. **한계**: Hidden representation 접근 필요(black-box 모델 불가), 3 지표의 layer·모델 architecture별 sensitivity 부담, RL 외 일반 post-training 일반화는 추가 검증.

🚀 실용적 활용

RL post-trained LLM의 평가 신뢰성 보장.
벤치마크 contamination 진단·정화.
Reasoning 모델 개발의 품질 관리.