Concept最佳-of-N采样 / Best-of-N
Best-of-N Sampling
An inference-time scaling technique that generates N candidate outputs in parallel for a given input and selects the best result using a scoring model. It improves the upper bound of generation quality without modifying the underlying model weights, serving as a straightforward method for test-time compute scaling.
Timeline (last 90 days)
Oct 12
The per-step redundancy mechanism has each sub-agent first plan and generate multiple candidate results, with its theoretical basis drawn from the Best-of-N sampling idea.
Unverified50%