用於推理的思維層級束搜尋

摘要

測試時計算擴展是大型推理模型（LRMs）效能提升的主要驅動力，但極度的低效率限制了現有方法，使得關鍵問題從「應投入多少計算」轉變為「應將計算配置於何處」。我們將測試時推理正式化為一個在部分軌跡上的受限計算配置問題。在固定的硬體預算下，現有典範無法主動將計算配置到最具潛力的部分進展上：傳統的並行抽樣將軌跡視為獨立處理，導致嚴重的記憶體瓶頸；而減除式剪枝則使硬體資源閒置，無法主動且充分地轉移輸出分佈。為了克服這種二分法，我們提出 Gambit，一種執行思路層級束搜尋的推論演算法。透過定期剪除缺乏前景的軌跡，並立即從高品質的前綴進行分支，Gambit 利用輕量級評分器探測隱藏狀態，動態地將計算集中於最具潛力的推理軌跡，同時維持持續的高硬體利用率。在多個模型與基準上的廣泛評估顯示，Gambit 嚴格優於現有基線方法。在相同的硬體限制下，我們的方法相較於剪枝基線，在 HMMT-24 上可獲得高達 +6.7% 的絕對準確率提升，在 AIME-25 上為 +3.3%；在軌跡完成上提供 >2 倍的吞吐量，並相較於標準並行抽樣，將總 token 消耗減少最多 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.