Concept
Best-of-N
An inference-time scaling strategy that independently samples N candidate outputs or model variants and selects the best according to a score. It carries a risk of overfitting due to selection bias.
Timeline (last 90 days)
Nov 3
The re-examination study found that in the best-of-N selection strategy, the top-expert reward improvement decays rapidly as the selection-set size m increases, indicating selection-set overfitting.
Unverified50%