This paper targets inefficient test-time compute scaling in large reasoning models, where independent sampling wastes resources and naive pruning fails to meaningfully shift the output distribution. It introduces Gambit, an inference algorithm that performs thought-level beam search, maintaining a fixed pool of reasoning traces while periodically pruning weak trajectories and branching new ones from high-quality prefixes. On math benchmarks (AIME, HMMT) and scientific reasoning tasks, Gambit improves accuracy by up to 6.7% over pruning baselines while cutting token consumption by 68.5% relative to standard parallel sampling.
