RADAR: Mechanistic Pathways for Detecting Data Contamination in LLM Evaluation

2025-10-09

AI Summary

Data contamination distorts LLM evaluation: models may achieve high scores by memorizing training data rather than performing genuine reasoning. This paper introduces RADAR, the first mechanism-based interpretability framework that distinguishes memory from reasoning behaviors in LLMs. RADAR extracts a 37-dimensional feature set—including surface-level confidence metrics and intrinsic mechanistic signals—from deep representational layers (e.g., attention specialization, circuit dynamics, and activation flow patterns) and employs an ensemble classifier for contamination detection. Evaluated across diverse test sets, RADAR achieves 93% overall accuracy, 100% accuracy on unambiguous cases, and 76.7% on ambiguous ones. Its core contributions are: (1) formalizing and quantifying the “memory–reasoning” behavioral mechanism in LLMs; (2) providing the first neuro-mechanism-grounded, interpretable detection framework specifically designed for assessment contamination; and (3) establishing a new benchmark and diagnostic toolkit for trustworthy LLM evaluation.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: Interpretability, Analysis, and Evaluation of NLP Models
• Knowledge Representation and Reasoning: Reasoning with Beliefs

Application Category

• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation

Abstract

Data contamination poses a significant challenge to reliable LLM evaluation, where models may achieve high performance by memorizing training data rather than demonstrating genuine reasoning capabilities. We introduce RADAR (Recall vs. Reasoning Detection through Activation Representation), a novel framework that leverages mechanistic interpretability to detect contamination by distinguishing recall-based from reasoning-based model responses. RADAR extracts 37 features spanning surface-level confidence trajectories and deep mechanistic properties including attention specialization, circuit dynamics, and activation flow patterns. Using an ensemble of classifiers trained on these features, RADAR achieves 93% accuracy on a diverse evaluation set, with perfect performance on clear cases and 76.7% accuracy on challenging ambiguous examples. This work demonstrates the potential of mechanistic interpretability for advancing LLM evaluation beyond traditional surface-level metrics.

Problem

• Detecting data contamination in LLM evaluation through mechanistic interpretability
• Distinguishing recall-based from reasoning-based model responses
• Advancing LLM evaluation beyond traditional surface-level metrics

Innovation

• Leverages mechanistic interpretability to detect data contamination
• Extracts surface-level confidence and deep mechanistic features
• Uses ensemble classifiers for high-accuracy contamination detection