Best-of-N sampling is a widely used inference-time scaling method for large language models: generate N candidates and select the one with the highest reward. But how this strategy behaves when the reward model is incomplete—missing some aspects of quality—has remained unclear. A new arXiv preprint addresses this gap by viewing Best-of-N not as a simple selection rule, but as a competition between the extreme reward values produced across the N samples.

The paper, titled "Tail Competition Explains Scaling in Best-of-N Sampling for Verifiable Problems," argues that the tail of the reward distribution drives the scaling behavior. Rather than focusing on average performance, the analysis centers on how the maximum reward among N samples competes with the imperfections of the reward model. This perspective offers a theoretical explanation for why and when Best-of-N scaling works, and where it may break down.

The abstract is brief, so the full details of the model and its implications are not yet available from this summary alone. However, the framing suggests that the effectiveness of Best-of-N depends on the shape of the reward distribution's tail and the degree of reward incompleteness. For verifiable problems—where correctness can be checked—this competition may behave differently than in open-ended tasks. The preprint is available on arXiv under identifier 2610.19362.