面向推理的思维级束搜索

摘要

测试时计算扩展是大型推理模型（LRMs）性能的主要驱动因素，但极度的低效限制了现有方法，并将关键问题从花费多少计算转变为将计算分配到哪里。我们将测试时推理形式化为在部分轨迹上的受限计算分配问题。在固定硬件预算下，现有范式无法主动将计算分配给最有前景的部分进展：传统的并行采样独立地处理轨迹，导致严重的内存瓶颈，而削减式剪枝则使硬件得不到充分利用，且无法主动且充分地改变输出分布。为克服这一两难困境，我们提出了 Gambit，一种执行思维级束搜索的推理算法。通过定期剪除无前途的轨迹并立即从高质量前缀进行分支，Gambit 通过轻量级评分器探测隐藏状态，动态地将计算集中到最有前景的推理轨迹上，同时保持持续的高硬件利用率。跨多个模型和基准的广泛评估表明，Gambit 严格优于现有基线。在相同硬件约束下，我们的方法相对于剪枝基线在 HMMT-24 上取得高达 +6.7% 的绝对准确率提升，在 AIME-25 上取得 +3.3% 的提升；将轨迹完成的吞吐量提高 2 倍以上，并且相对于标准并行采样，总词元消耗最多减少 68.5%。

English

Test-time compute scaling is a primary driver of performance in large reasoning models (LRMs), but extreme inefficiency bounds current approaches, shifting the critical question from how much compute to spend, to where to allocate it. We formalize test-time reasoning as a constrained compute allocation problem over partial trajectories. Under a fixed hardware budget, existing paradigms fail to actively allocate the compute to the most promising partial progress: traditional parallel sampling treats traces independently and induces severe memory bottlenecks, while subtractive pruning starves hardware and fails to actively and sufficiently shift the output distribution. To overcome this dichotomy, we introduce Gambit, an inference algorithm that executes thought-level beam search. By periodically pruning unpromising trajectories and immediately branching from high-quality prefixes, Gambit dynamically concentrates compute onto the most promising reasoning traces via a light-weight scorer probing hidden states while maintaining continuous high hardware utilization. Extensive evaluations across multiple models and benchmarks demonstrate that Gambit strictly dominates existing baselines. Under identical hardware constraints, our method yields up to a +6.7\% absolute accuracy gain on HMMT-24 and +3.3\% on AIME-25 over pruning baselines, delivers >2times higher throughput on trace completion, and reduces total token consumption by up to 68.5\% relative to standard parallel sampling.