Gambit Boosts Reasoning Model Performance with Beam Search.

Scaling compute during test-time is a primary factor in the performance of large reasoning models (LRMs), but current approaches are often inefficient, leading to questions about optimal compute allocation. Existing paradigms, such as parallel sampling or subtractive pruning, fail to actively direct compute to the most promising partial progress, resulting in memory bottlenecks or underutilized hardware. This research introduces Gambit, an innovative inference algorithm designed to overcome these limitations by executing "thought-level beam search." Gambit dynamically concentrates computational resources on the most promising reasoning paths. It achieves this by periodically identifying and pruning less promising trajectories, while simultaneously branching from high-quality prefixes. This process is guided by a lightweight scorer that probes hidden states, ensuring continuous high hardware utilization. Extensive evaluations across various models and benchmarks demonstrate that Gambit consistently outperforms existing baselines. Under identical hardware constraints, it delivers substantial accuracy gains (e.g., +6.7% on HMMT-24), more than doubles throughput on trace completion, and significantly reduces total token consumption by up to 68.5% compared to standard parallel sampling.