AI evaluation arenas โ platforms where models compete head-to-head and accumulate Elo-like ratings โ have a strategic vulnerability. A model producer can submit multiple variants of the same model, creating clones that inflate the original's ranking through correlated performance. If clone A and clone B both perform similarly, their pairwise comparisons don't generate useful information but their interactions with other models create rating artifacts.
Hays, Li, Flanigan, and Raghavan formalize when cloning becomes profitable and propose a defense: You-Rank-We-Rank (YRWR). The mechanism requires each producer to submit a self-ranking of their own models. If you submit five variants, you must declare which you think is best, second-best, and so on. The system then uses these self-rankings to correct the quality estimates, deflating the artificial boost from correlated submissions.
The elegant property: YRWR is approximately clone-robust even when producers misrank their own models. The self-ranking doesn't need to be honest to be useful โ it provides structural information about which submissions are variants of each other, which is enough to correct the statistical distortion.
The broader pattern: evaluation systems that assume independent submissions fail when submissions are strategically correlated. The fix isn't preventing correlation (you can't stop people from submitting variants) but accounting for it structurally. The self-ranking requirement doesn't punish cloning โ it neutralizes its advantage. The strategic incentive to clone disappears not because cloning is banned but because the evaluation absorbs the correlation. The defense works by making the attack unprofitable, not by making it impossible.