摆脱冗余推理:面向推理阶段大语言模型的结构感知搜索
文章背景与核心概要
在使用大语言模型(LLM)进行推理阶段搜索时,往往会面临“推理盆地坍缩”(reasoning basin collapse)的问题。这是一种失效模式,即搜索过程过度集中于一小组结构或语义相似的轨迹上,导致其他可行替代方案未得到充分探索。
为了克服这一难题,本文引入了 BASIN 这一无需训练的结构感知选择方法。BASIN 将推理状态划分为不同的“盆地”(basins),并对重复访问相同策略的行为进行惩罚,从而确保在固定的计算预算下,搜索资源能够重新分配到真正截然不同的推理路径上。其质量感知变体 QA-BASIN 则通过保留高质量的推理盆地,进一步增强了鲁棒性。此外,作者还引入了冗余差距(\(\Delta\))来量化正确预测与错误预测之间搜索集中度的差异,并证明了 BASIN 如何成功将 \(\Delta\) 转化为正值。
摘要 (Abstract)
Inference-time search with large language models (LLMs) often concentrates on a small set of structurally or semantically similar trajectories, leaving alternatives underexplored—a failure mode we call reasoning basin collapse. We introduce BASIN, a training-free, structure-aware selection method that groups reasoning states into basins and penalizes repeated visits to the same strategy, thereby reallocating search across genuinely distinct reasoning paths under a fixed compute budget. Under matched inference budgets, BASIN improves over Tree of Thoughts (ToT) by up to \(+22\)pp on Game of 24 and \(+6.7\)pp on MuSR. A quality-aware variant, QA-BASIN, further improves robustness by preserving high-quality basins when unconditional diversification over-explores. To explain when basin-aware selection helps, we introduce the redundancy gap \(\Delta\), which measures how differently search concentrates for correct versus incorrect predictions: standard ToT often operates near \(\Delta \approx 0\), while BASIN consistently shifts \(\Delta\) positive. More broadly, BASIN suggests structure-aware selection as a simple and general approach to improving inference-time reasoning.
使用大语言模型(LLM)进行的推理阶段搜索常常集中于一小组合成或语义相似的轨迹,使得其他替代方案未被充分探索——我们将这种失效模式称为推理盆地坍缩(reasoning basin collapse)。为此,我们引入了 BASIN,这是一种无需训练的结构感知选择方法,它将推理状态分组为盆地,并对重复访问相同策略的行为进行惩罚,从而在固定的计算预算下,将搜索重新分配到真正独特的推理路径上。在匹配的推理预算下,在 Game of 24 基准测试中,BASIN 相比思维树(ToT)提升高达 \(+22\)个百分点(pp),在 MuSR 上提升高达 \(+6.7\)pp。其质量感知变体 QA-BASIN 在无条件多样化过度探索时,通过保留高质量盆地进一步增强了鲁棒性。为了解释盆地感知选择在何时发挥作用,我们引入了冗余差距 \(\Delta\),它用于衡量正确预测与错误预测在搜索集中度上的差异:标准的 ToT 通常运行在 \(\Delta \approx 0\) 附近,而 BASIN 则持续将 \(\Delta\) 转化为正值。从更广泛的意义上讲,BASIN 表明结构感知选择是一种改善推理阶段推理的简单且通用的方法。
元数据 (Metadata)
- arXiv ID: arXiv:2609.00738 [cs.AI]
- Subject: Artificial Intelligence (
cs.AI)- Author: Lu Cheng
- Submission Date: September 1, 2026
- License: Creative Commons Attribution 4.0 International
- arXiv ID: arXiv:2609.00738 [cs.AI]
- 学科: 人工智能 (
cs.AI) - 作者: Lu Cheng
- 提交日期: 2026年9月1日
- 许可协议: 知识共享署名 4.0 国际许可协议
核心贡献与结果 (Key Contributions & Results)
- BASIN Algorithm: A novel, training-free, structure-aware selection mechanism designed to prevent reasoning basin collapse during inference-time LLM search.
- QA-BASIN: An extension of the method that is quality-aware, preventing over-exploration by safeguarding high-performing reasoning paths.
- Empirical Gains:
- Up to \(+22\text{pp}\) improvement over Tree of Thoughts (ToT) on the Game of 24 benchmark.
- Up to \(+6.7\text{pp}\) improvement on MuSR.
- The Redundancy Gap (\(\Delta\)): A proposed metric analyzing search concentration behavior for correct versus incorrect model outputs.
- BASIN 算法: 一种新颖的、无需训练的结构感知选择机制,旨在防止 LLM 推理阶段搜索过程中的推理盆地坍缩。
- QA-BASIN: 该方法的质量感知扩展,通过保护高性能的推理路径来防止过度探索。
- 实验增益:
- 在 Game of 24 基准测试中,相比思维树(ToT)提升高达 \(+22\text{pp}\)。
- 在 MuSR 上提升高达 \(+6.7\text{pp}\)。
- 冗余差距(\(\Delta\)): 提出了一种分析模型正确输出与错误输出搜索集中度行为的指标。
资源与链接 (Resources & Links)
- PDF Version: View PDF
- HTML Version: arXiv HTML (Experimental)
- Source Code: GitHub Repository
- DOI: 10.48550/arXiv.2609.00738
- PDF 版本: 查看 PDF
- HTML 版本: arXiv HTML(实验性)
- 源代码: GitHub 仓库
- DOI: 10.48550/arXiv.2609.00738