Leaderboard Hacking: Preference-Based Model Evaluations are Vulnerable to Manipulation

Summary

A new arXiv study (2611.14769) shows that arena-style leaderboards using the Bradley-Terry model are vulnerable to manipulation. Strategies like strategic voting, prompting, and nomination can boost a model's rank by up to 15-20 places individually, and up to 45 places when combined. The authors demonstrate these effects on public data and propose safeguards.

This is a reviewed summary. Read the complete article at the original source.

Why it matters by industry

• Artificial intelligence

  This is directly relevant to AI model evaluation and ranking, a core aspect of model development and deployment. The findings highlight security vulnerabilities in common evaluation frameworks, impacting trust in leaderboards and model selection.