Why it matters

Search scaling extends inference-time compute. Understanding shapes design.

Advertisement

The architecture

Sample many candidates.

Evaluate via verifier.

Expand promising; prune others.

Search-based inference scalingSample candidatesN per stepEvaluateverifier / valueExpand + pruneMCTS / beamAlphaProof + o1 + K1.5 use search variants; verifier quality critical
Search scaling.
Advertisement

How it works end to end

MCTS for game-like reasoning.

Beam search for text.

Best-of-N simpler baseline.