This research reveals preference leakage - a critical contamination issue that occurs when LLMs are used to judge outputs from other AI systems trained on synthetic data.

• LLMs serving as judges can show bias toward outputs created by models similar to themselves
• This contamination significantly inflates performance metrics, creating a false sense of model capability
• Researchers found evidence of preference leakage across multiple popular LLM architectures
• Detection methods were developed to identify and mitigate this previously overlooked security vulnerability

For security professionals, this research exposes a fundamental flaw in current AI evaluation practices that could lead to deploying models with overestimated capabilities and unknown risks.

Preference Leakage: A Contamination Problem in LLM-as-a-judge