跳转至

文章背景与核心概要

大型推理模型（LRMs）在针对同一问题进行多次查询时，往往会生成多样化甚至相互矛盾的答案。尽管在 \(k\) 次推演（rollouts）中采用多样本多数表决是确保可靠部署的黄金标准，但在大模型所需的规模下，这种方法的计算成本极其高昂。

本文引入了一种创新的推理时方法——思维漏斗（Funnel of Thoughts, FoT）。该方法在保持 32 条轨迹投票完整准确率的同时，将注意力机制的 FLOPs（浮点运算次数）减半（实现了全模型推理成本 28.8% 的总体降低）。通过对来自 6 个大型推理模型的 11.5 万条推理轨迹进行分析，作者发现非生产性的轨迹往往会通过重复的犹豫标记（例如 "Wait"、"Actually" 和 "perhaps"）暴露出来。这些思维循环极少能引导至正确答案，却消耗了大量的计算资源。

利用这种无需训练的词汇信号，FoT 能够在生成中期识别并剪枝失败的轨迹，将在线生成的注意力 FLOPs 降低了 56.1%，壁钟时间（wall-clock time）缩短了 37.6%，且无需针对不同架构和领域进行额外的模型推理或重新训练。

Funnel of Thoughts: Efficient Test-Time Scaling via Early Voting and Rollout Pruning

arXiv: 2608.15065 [cs.AI]
DOI: 10.48550/arXiv.2608.15065
Submitted: August 15, 2026
Authors: Chanhee Park, Sungbin Han, Jeongho Yoon, Seongtae Hong, Heuiseok Lim

Funnel of Thoughts: Efficient Test-Time Scaling via Early Voting and Rollout Pruning

arXiv: 2608.15065 [cs.AI]
DOI: 10.48550/arXiv.2608.15065
Submitted: August 15, 2026
Authors: Chanhee Park, Sungbin Han, Jeongho Yoon, Seongtae Hong, Heuiseok Lim

📌 Summary

Large Reasoning Models (LRMs) often generate diverse and sometimes contradictory answers across multiple queries for the same problem. While multi-sample majority voting at \(k\) rollouts is the gold standard for reliable deployment, it is computationally prohibitive at the scale required by LRMs.

This paper introduces Funnel of Thoughts (FoT), an innovative inference-time method that maintains the full accuracy of 32-trajectory voting while cutting attention FLOPs in half (achieving a 28.8% overall reduction in full-model inference cost). By analyzing 115K reasoning trajectories from six LRMs, the authors discovered that unproductive trajectories often expose themselves through repeated hesitation markers (e.g., "Wait", "Actually", and "perhaps"). These loops rarely lead to the correct answer while consuming massive computational resources.

Using this training-free lexical signal, FoT spots and prunes failing trajectories mid-generation, reducing online generation attention FLOPs by 56.1% and wall-clock time by 37.6% without requiring additional model inference or retraining across different architectures and domains.

📌 Summary

Large Reasoning Models (LRMs) often generate diverse and sometimes contradictory answers across multiple queries for the same problem. While multi-sample majority voting at \(k\) rollouts is the gold standard for reliable deployment, it is computationally prohibitive at the scale required by LRMs.

This paper introduces Funnel of Thoughts (FoT), an innovative inference-time method that maintains the full accuracy of 32-trajectory voting while cutting attention FLOPs in half (achieving a 28.8% overall reduction in full-model inference cost). By analyzing 115K reasoning trajectories from six LRMs, the authors discovered that unproductive trajectories often expose themselves through repeated hesitation markers (e.g., "Wait", "Actually", and "perhaps"). These loops rarely lead to the correct answer while consuming massive computational resources.

Using this training-free lexical signal, FoT spots and prunes failing trajectories mid-generation, reducing online generation attention FLOPs by 56.1% and wall-clock time by 37.6% without requiring additional model inference or retraining across different architectures and domains.

📋 Paper Metadata

📋 Paper Metadata