Preference Leakage: A Contamination Problem in LLM-as-a-judge
This addresses a contamination issue in LLM-driven data annotation for model development, which is incremental as it builds on known biases in LLM-as-a-judge methods.
The paper identifies preference leakage, a contamination problem in LLM-as-a-judge scenarios where synthetic data generators and evaluators are related, causing bias in judgments. Through experiments across multiple LLM baselines and benchmarks, it empirically confirms this bias and shows it is pervasive and hard to detect.
Large Language Models (LLMs) as judges and LLM-based data synthesis have emerged as two fundamental LLM-driven data annotation methods in model development. While their combination significantly enhances the efficiency of model training and evaluation, little attention has been given to the potential contamination brought by this new model development paradigm. In this work, we expose preference leakage, a contamination problem in LLM-as-a-judge caused by the relatedness between the synthetic data generators and LLM-based evaluators. To study this issue, we first define three common relatednesses between the data generator LLM and the judge LLM: being the same model, having an inheritance relationship, and belonging to the same model family. Through extensive experiments, we empirically confirm the bias of judges towards their related student models caused by preference leakage across multiple LLM baselines and benchmarks. Further analysis suggests that preference leakage is a pervasive and real-world problem that is harder to detect compared to previously identified biases in LLM-as-a-judge scenarios. All of these findings imply that preference leakage is a widespread and challenging problem in the area of LLM-as-a-judge. We release all codes and data at: https://github.com/David-Li0406/Preference-Leakage.