A paper posted to arXiv on 10 Nov 2026 reveals that arena‑style leaderboards for language models are easy to game. The authors demonstrate three attacks – strategic voting, prompting, and nomination – that can each raise a model’s rank by about 15‑20 places. When combined, the tricks can move a model up roughly 45 spots on the leaderboard. The paper also outlines safeguards, but notes the problem applies to any system using the Bradley‑Terry model for pairwise preferences.
Why it matters
If unchecked, these hacks could mislead developers and users about which language model truly performs best.
