Reproduce “Tail Competition Explains Scaling in Best-of-N Sampling for Verifiable Problems” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Taro Yano, Yoichi Ishibashi, Masafumi Oyamada
The authors analyze Best-of-N sampling on verifiable tasks, modeling selection as a contest between maximum scores for correct and incorrect responses. Under Gaussian scores, higher incorrect-response variance can produce early gains followed by declining accuracy as sampling grows. Their stopping method improves accuracy while generating fewer samples on the tested reward-hacking subset.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.