“Researchers at frontier AI labs feel pressured to optimize models for leaderboard performance, such as on LM Arena, for career advancement, even when they believe it degrades the model's accuracy.”