Best-of-N Sampling

A test-time strategy that generates N candidate outputs and selects the single best via a verifier, judge, or gate rather than accepting the first sample.

grounded in: idea_engine.py proposes up to 3 candidate concepts and a downstream gate / self_improve selects the single highest-value change ('pick the single highest-value change from real signals'), a generate-N

Connected concepts

Self-Consistency, LLM-as-judge, Test-time compute, Idea engine, Verification loops