Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

RADAR: Mechanistic Pathways for Detecting Data Contamination in LLM Evaluation

Ashish Kattamuri, Harshwardhan Fartale, Arpita Vats, Rahul Raja, Ishita Prasad

TL;DR

Data contamination undermines reliable LLM evaluation by conflating memorization with genuine reasoning. RADAR uses mechanistic interpretability to distinguish recall from reasoning by extracting 37 surface and internal features and employing an ensemble of four classifiers, achieving about $93\%$ overall accuracy and strong performance on clear cases while handling challenging prompts. The approach highlights interpretable signatures such as specialized attention and circuit dynamics, and operates without access to training data, offering a practical tool to improve evaluation reliability. This work suggests a scalable direction for LLM assessment that complements traditional metrics by examining internal computation rather than solely outputs.

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.

RADAR: Mechanistic Pathways for Detecting Data Contamination in LLM Evaluation

TL;DR

Data contamination undermines reliable LLM evaluation by conflating memorization with genuine reasoning. RADAR uses mechanistic interpretability to distinguish recall from reasoning by extracting 37 surface and internal features and employing an ensemble of four classifiers, achieving about overall accuracy and strong performance on clear cases while handling challenging prompts. The approach highlights interpretable signatures such as specialized attention and circuit dynamics, and operates without access to training data, offering a practical tool to improve evaluation reliability. This work suggests a scalable direction for LLM assessment that complements traditional metrics by examining internal computation rather than solely outputs.

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.

Paper Structure

This paper contains 39 sections, 4 equations, 2 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Methodology
  3. Framework Architecture
  4. Feature Engineering
  5. Classification System
  6. Experiments and Results
  7. Experimental Setup and Results
  8. Feature Analysis
  9. Discussion and Implications
  10. Contamination Detection Applications
  11. Interpretability Insights
  12. Conclusion
  13. Implementation Details
  14. Model Configuration
  15. Feature Computation
  16. ...and 24 more sections

Figures (2)

  • Figure 1: RADAR Framework Architecture: Input prompts are processed by the Mechanistic Analyzer to extract internal states, which are converted to Surface and Mechanistic Features, then classified by an ensemble to predict recall vs. reasoning with confidence scores.
  • Figure 2: RADAR Feature Analysis: Comparison of surface and mechanistic features for recall and reasoning tasks, highlighting top discriminative features and RDS–RCI score distribution. The results show recall tasks characterized by early confidence and specialized heads, while reasoning tasks rely on broader circuit complexity and higher activation flow variance. The scatter plot demonstrates strong clustering, with recall tasks in the high-RDS region and reasoning tasks distributed in lower-RDS regions.