Best-of-N Sampling Test: Math Accuracy Boosted to 65%

SergioPaniego · x · 2026-08-23

Hugging Face shared a test report on an agent using Best-of-N sampling with a reward model to solve math problems. The experiment used Qwen2.5-1.5B as the generator and Skywork PRM as the scorer. Results show that Weighted Best-of-N (N=16) achieved 65% accuracy on the MATH dataset, significantly outperforming greedy decoding at 45%. The method samples multiple solutions, groups and scores them, and selects the answer with the highest weighted score, effectively rescuing correct answers that appear infrequently.

Original post →

More from Research

Research channel →