mlsanityData to 7 Oct 2026

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

arXiv 2605.29888 ↗First posted 28 May 2026cs.LG cs.AIcited by 0

On the map: LLM reasoning → Reinforcement learning for reasoning

Abstract (arXiv)

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.

Key facts. No citing papers in the map yet (data through 7 Oct 2026). Placed in Reinforcement learning for reasoning, one of 17 directions in LLM reasoning. Builds on 29 of its 43 references that are in the map. First posted on arXiv 28 May 2026.

Builds on

29 of 43 references.

Closest ancestors · 4

Same direction, newest first

  1. 2025Reasoning or Memorization? Unreliable Results of Reinforcement Learning Due to Data Contamination14 Jul
  2. 2025Learning to Reason without External Rewards26 May
  3. 2025LIMR: Less is More for RL Scaling17 Feb
  4. 2025Process Reinforcement through Implicit Rewards3 Feb

Other references · 25

  1. 2023Efficient Memory Management for Large Language Model Serving with PagedAttentioncited by 5,767
  2. 2024HybridFlow: A Flexible and Efficient RLHF Frameworkcited by 1,980
  3. 2023Inference-Time Intervention: Eliciting Truthful Answers from a Language Modelcited by 1,259
  4. 2023Steering Language Models With Activation Engineeringcited by 824
  5. 2024Training Large Language Models to Reason in a Continuous Latent Spacecited by 763
  6. 2025OpenThoughts: Data Recipes for Reasoning Modelscited by 287
  7. 2025MathArena: Evaluating LLMs on Uncontaminated Math Competitionscited by 275
  8. 2022Demystifying Prompts in Language Models via Perplexity Estimationcited by 215
  9. 2024Programming Refusal with Conditional Activation Steeringcited by 191
  10. 2025Scalable Best-of-N Selection for Large Language Models via Self-Certaintycited by 182
  11. 2023Time Travel in LLMs: Tracing Data Contamination in Large Language Modelscited by 175
  12. 2024Taming Overconfidence in LLMs: Reward Calibration in RLHFcited by 105
13 more
  1. 2025A Sober Look at Progress in Language Model Reasoning: Pitfalls and Paths to Reproducibilitycited by 89
  2. 2024ReCaLL: Membership Inference via Relative Conditional Log-Likelihoodscited by 61
  3. 2024Latent Space Chain-of-Embedding Enables Output-free LLM Self-Evaluationcited by 56
  4. 2025Tracing the Representation Geometry of Language Models from Pretraining to Post-trainingcited by 33
  5. 2025Restoring Calibration for Aligned Large Language Models: A Calibration-Aware Fine-Tuning Approachcited by 25
  6. 2026Manifold Steering Reveals the Shared Geometry of Neural Network Representation and Behaviorcited by 22
  7. 2025Membership Inference Attacks on Large-Scale Models: A Surveycited by 17
  8. 2025How Contaminated Is Your Benchmark? Quantifying Dataset Leakage in Large Language Models with Kernel Divergencecited by 13
  9. 2025On The Fragility of Benchmark Contamination Detection in Reasoning Modelscited by 8
  10. 2026G-Drift MIA: Membership Inference via Gradient-Induced Feature Drift in LLMscited by 6
  11. 2025Training-free LLM Verification via Recycling Few-shot Examplescited by 4
  12. 2025Neural Breadcrumbs: Membership Inference Attacks on LLMs Through Hidden State and Attention Pattern Analysiscited by 3
  13. 2026Gap-K%: Measuring Top-1 Prediction Gap for Detecting Pretraining Datacited by 1

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2025Detecting Data Contamination from Reinforcement Learning Post-training for Large Language Models
  2. 2026The Impact of Post-training on Data Contamination
  3. 2026Detecting RLVR Training Data via Structural Convergence of Reasoning
  4. 2026Noisy Data is Destructive to Reinforcement Learning with Verifiable Rewards
  5. 2025Reshaping Reasoning in LLMs: A Theoretical Analysis of RL Training Dynamics through Pattern Selection
  6. 2025Quagmires in SFT-RL Post-Training: When High SFT Scores Mislead and What to Use Instead
  7. 2025On the Interplay of Pre-Training, Mid-Training, and RL on Reasoning Language Models
  8. 2026Sparse but Critical: A Token-Level Analysis of Distributional Shifts in RLVR Fine-Tuning of LLMs