Gambit：思想级束搜索提升推理模型效率

AI 摘要

Gambit是一种新的推理算法，通过思想级束搜索在固定硬件预算下动态分配计算资源到有前景的推理轨迹。在多个模型和基准测试中，Gambit相比现有基线在HMMT-24上取得+6.7%的绝对准确率提升，在AIME-25上提升+3.3%，同时将总token消耗减少高达68.5%。该方法通过轻量级评分器探测隐藏状态，定期剪枝无望轨迹并立即从高质量前缀分支，保持高硬件利用率。

正文节选

Thought-Level Beam Search for Reasoning Abstract Gambit improves reasoning model efficiency by using thought-level beam search to dynamically allocate compute to promising reasoning traces under fixed hardware budgets. 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 constrai

发布时间：—

抓取时间：2026-08-15 07:37

来源机构：Hugging Face