LaRA: Layer-wise Representation Analysis for Detecting Data Contamination in RL Post-Training
Summary
LaRA is a layer-wise representation analysis framework that detects data contamination in RL post-trained LLMs by measuring geometric deviations across model layers, outperforming output-level baselines.
View Cached Full Text
Cached at: 05/29/26, 03:00 AM
Paper page - LaRA: Layer-wise Representation Analysis for Detecting Data Contamination in RL Post-Training
Source: https://huggingface.co/papers/2605.29888
Abstract
LaRA is a layer-wise representation analysis framework that detects data contamination in reinforcement learning-post-trained large language models by analyzing geometric deviations across model layers.
Reinforcement learning(RL) post-training has shown to improve reasoning inlarge language models(LLMs). However, there has been little exploration on the problem ofdata contaminationin 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, alayer-wise representation analysisframework for detecting contamination in RL post-trained LLMs. LaRA introduces three complementary metrics, measuringperturbation sensitivity,directional collapse, andlocal representation rigidityunder controlled perturbations. We find that contamination produces progressivegeometric deviationsacross layers, including amplifiedperturbation sensitivity, strongerdirectional 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.
View arXiv pageView PDFAdd to collection
Get this paper in your agent:
hf papers read 2605\.29888
Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash
Models citing this paper0
No model linking this paper
Cite arxiv.org/abs/2605.29888 in a model README.md to link it from this page.
Datasets citing this paper0
No dataset linking this paper
Cite arxiv.org/abs/2605.29888 in a dataset README.md to link it from this page.
Spaces citing this paper0
No Space linking this paper
Cite arxiv.org/abs/2605.29888 in a Space README.md to link it from this page.
Collections including this paper0
No Collection including this paper
Add this paper to acollectionto link it from this page.
Similar Articles
Layer-wise Representation Dynamics: An Empirical Investigation Across Embedders and Base LLMs
This paper introduces Layer-wise Representation Dynamics (LRD), a framework with three measurement families to analyze how hidden states change across layers in language models. Applied to 31 models on 30 MTEB tasks, LRD reveals architectural differences and enables label-free model selection and inference-time layer pruning.
RDP LoRA: Geometry-Driven Identification for Parameter-Efficient Adaptation in Large Language Models
RDP-LoRA uses geometric trajectory analysis and the Ramer-Douglas-Peucker algorithm to automatically select the most impactful layers for parameter-efficient fine-tuning, outperforming full-layer and random LoRA baselines.
Hybrid-LoRA: Bridging Full Fine-Tuning and Low-Rank Adaptation for Post-Training
Hybrid-LoRA proposes a framework that selectively applies full fine-tuning to a small subset of modules while using LoRA for the rest, achieving performance near full fine-tuning with significantly lower computational cost. Experiments show improvements of up to 5.65% over existing parameter-efficient baselines.
A Training-Time Diagnostic for Generalization via the Log-Alignment Ratio
This paper introduces the log-alignment ratio (LAR), a training-time metric that measures parameter-activation alignment and predicts generalization by capturing the spread of weight and activation spectra. Experiments on grokking and a 3B-parameter language model show LAR tracks the transition from memorization to generalization and flags overfitting without held-out data.
DART: Mitigating Harm Drift in Difference-Aware LLMs via Distill-Audit-Repair Training
DART (Distill-Audit-Repair Training) is a new training framework that addresses 'harm drift' in safety-aligned LLMs, where fine-tuning for demographic difference-awareness causes harmful content to appear in model explanations. On eight benchmarks, DART improves Llama-3-8B-Instruct accuracy from 39.0% to 68.8% while reducing harm drift cases by 72.6%.