Detecting Distillation Data from Reasoning Models

2025-10-06

AI Summary

To address benchmark contamination and inflated performance caused by evaluation-data leakage in inference-model distillation, this paper proposes Token Probability Deviation (TBD), a detection method that quantifies the deviation between generated token probabilities and high-quality reference probabilities to identify whether distilled data has been exposed to evaluation-set questions—enabling efficient detection even under partial data availability. TBD models the output probability distribution of language models and requires neither access to original training data nor knowledge of distillation process details. On the S1 dataset, TBD achieves an AUC of 0.918 and a TPR@1%FPR of 0.470, substantially outperforming existing baselines. This work introduces the first probabilistic-deviation-based approach for distillation-data provenance tracing, establishing a scalable, low-overhead detection paradigm to mitigate benchmark pollution.

Technology Category

• Machine Learning: Calibration & Uncertainty Quantification
• Natural Language Processing: Interpretability, Analysis, and Evaluation of NLP Models
• Reasoning under Uncertainty: Relational Probabilistic Models

Application Category

• Web Mining and Content Analysis: Web data provenance, reliability, and authenticity
• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

Reasoning distillation has emerged as an efficient and powerful paradigm for enhancing the reasoning capabilities of large language models. However, reasoning distillation may inadvertently cause benchmark contamination, where evaluation data included in distillation datasets can inflate performance metrics of distilled models. In this work, we formally define the task of distillation data detection, which is uniquely challenging due to the partial availability of distillation data. Then, we propose a novel and effective method Token Probability Deviation (TBD), which leverages the probability patterns of the generated output tokens. Our method is motivated by the analysis that distilled models tend to generate near-deterministic tokens for seen questions, while producing more low-probability tokens for unseen questions. Our key idea behind TBD is to quantify how far the generated tokens' probabilities deviate from a high reference probability. In effect, our method achieves competitive detection performance by producing lower scores for seen questions than for unseen questions. Extensive experiments demonstrate the effectiveness of our method, achieving an AUC of 0.918 and a TPR@1% FPR of 0.470 on the S1 dataset.

Problem

• Detecting distillation data to prevent benchmark contamination
• Identifying seen versus unseen questions in reasoning models
• Quantifying token probability deviations for contamination detection

Innovation

• Detects distillation data via token probability deviation
• Quantifies deviation from high reference probability
• Distinguishes seen questions using lower deviation scores