Thought-Level Beam Search for Reasoning
Gambit reallocates inference compute mid-reasoning by pruning weak traces and branching from stronger partial thoughts.
The paper frames test-time reasoning as a fixed-budget compute allocation problem. Its method, thought-level beam search, uses a lightweight hidden-state scorer to decide which partial trajectories continue and which prefixes get expanded. In the authors’ evaluations, Gambit beats pruning and parallel sampling baselines under the same hardware limits, with reported gains including +6.7% on HMMT-24 and +3.3% on AIME-25. It also reports more than 2x higher trace-completion throughput and up to 68.5% fewer tokens than standard parallel sampling. HF Daily Papers' note
The paper frames test-time reasoning as a fixed-budget compute allocation problem. Its method, thought-level beam search, uses a lightweight hidden-state scorer to decide which partial trajectories continue and which prefixes get expanded. In the authors’ evaluations, Gambit beats pruning and parallel sampling baselines under the same hardware limits, with reported gains including +6.7% on HMMT-24 and +3.3% on AIME-25. It also reports more than 2x higher trace-completion throughput and up to 68.5% fewer tokens than standard parallel sampling. HF Daily Papers' note
score 5