PaCoST: Paired Confidence Significance Testing for Benchmark Contamination Detection in Large Language Models

TL;DR

This work addresses benchmark contamination in LLM evaluation by introducing PaCoST, a threshold-free method that compares model confidence on original versus paraphrased inputs. It uses a three-step pipeline—rephrase preparation, confidence estimation with $P(True)$, and a paired-sample $t$-test with a significance level of $p<0.05$—to detect contamination without requiring access to training data. The authors formalize contamination objectives and derive five robust detection requirements, demonstrating PaCoST's effectiveness and stability across multiple open-source models and benchmarks. Their extensive experiments reveal pervasive suspected contamination across many benchmarks and models, underscoring the need for benchmark-free or dynamic evaluation strategies and independent auditing tools to maintain trust in LLM benchmarking.

Abstract

Large language models (LLMs) are known to be trained on vast amounts of data, which may unintentionally or intentionally include data from commonly used benchmarks. This inclusion can lead to cheatingly high scores on model leaderboards, yet result in disappointing performance in real-world applications. To address this benchmark contamination problem, we first propose a set of requirements that practical contamination detection methods should follow. Following these proposed requirements, we introduce PaCoST, a Paired Confidence Significance Testing to effectively detect benchmark contamination in LLMs. Our method constructs a counterpart for each piece of data with the same distribution, and performs statistical analysis of the corresponding confidence to test whether the model is significantly more confident under the original benchmark. We validate the effectiveness of PaCoST and apply it on popular open-source models and benchmarks. We find that almost all models and benchmarks we tested are suspected contaminated more or less. We finally call for new LLM evaluation methods.

Figures (1)

• Figure 1: Overview of our method. $x_i$ represents a question, $y_i$ represents its corresponding ground truth answer, $x_i^{'}$ represents a rephrased question and $a_{i}, a_{i}^{'}$ represent model responses to original and rephrased question correspondingly.