←Back to feed
🧠 AI🔴 Bearish
Preference Leakage: A Contamination Problem in LLM-as-a-judge
arXiv – CS AI|Dawei Li, Renliang Sun, Yue Huang, Ming Zhong, Bohan Jiang, Jiawei Han, Xiangliang Zhang, Wei Wang, Huan Liu|
🤖AI Summary
Researchers have identified 'preference leakage,' a contamination problem in LLM-as-a-judge systems where evaluator models show bias toward related data generator models. The study found this bias occurs when judge and generator LLMs share relationships like being the same model, having inheritance connections, or belonging to the same model family.
Key Takeaways
- →Preference leakage causes LLM judges to unfairly favor outputs from related generator models during evaluation.
- →The bias occurs across three relationship types: same model, inheritance relationship, and same model family.
- →This contamination problem is pervasive across multiple LLM baselines and benchmarks in real-world scenarios.
- →Preference leakage is harder to detect compared to previously identified biases in LLM evaluation systems.
- →The research provides open-source code and data to help the community address this evaluation integrity issue.
#llm#artificial-intelligence#bias#evaluation#contamination#model-training#data-synthesis#research#arxiv
Read Original →via arXiv – CS AI
Act on this with AI
Stay ahead of the market.
Connect your wallet to an AI agent. It reads balances, proposes swaps and bridges across 15 chains — you keep full control of your keys.
Related Articles