arXiv cs.AI· Lingyu Jiang, Zirui Li, Lingyu Xu, Shuo Xing, Peiran Li, Tsubasa Takahashi, Dengzhe Hou, Zhengzhong Tu, Kazunori Yamada, Fangzhou Lin·· 2 天前AI 评分22
PathCal：面向高效推理的状态感知反思标记校准方法
PathCal: State-Aware Reflection-Marker Calibration for Efficient Reasoning
AI 导读
PathCal 是一种免训练控制器，通过延续、修正与替代开启三类反思标记的瞬时竞争，动态调整固定延续偏置的强度，且无需额外前向传播。实验显示，PathCal 在多种数学推理场景中改善了准确率与长度的权衡，并在代码与多语言数学任务中缩短了生成长度。
正文 · 原文
Abstract:Large Reasoning Language Models (LRMs) tackle complex tasks through test-time scaling, generating long-form Chain-of-Thought (CoT) trajectories during inference. These trajectories contain reflection markers such as wait, but, and alternatively, accompanying hesitation, revision, and alternative explorations. Recent test-time control methods use these markers to steer reasoning, but treating them as one coarse-grained category can obscure their distinct intervention effects. We study how marker identity and reasoning context shape these effects. Type-wise suppression produces different accuracy--length tradeoffs, while fixed-prefix interventions show that the relative benefit of continuation and revision varies across prefixes and tasks. Motivated by these findings, we introduce PathCal, a training-free controller that adjusts the strength of a fixed continuation bias using instantaneous competition among continuation, revision, and alternative-opening markers. It applies finite logit adjustments without additional model forward passes. An exact log-odds identity and frozen-prefix probes establish its local directional effect. Experiments show that PathCal improves the accuracy--length tradeoff in many mathematical reasoning settings and also reduces generation length in coding and multilingual mathematics.
| Comments: | 21 pages, 5 figures, 24 tables |
| Subjects: | Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2605.23074 [cs.AI] |
| (or arXiv:2605.23074v2 [cs.AI] for this version) | |
| https://doi.org/10.48550/arXiv.2605.23074 arXiv-issued DOI via DataCite |
Submission history
From: Fangzhou Lin [view email]
[v1]
Thu, 21 May 2026 22:13:20 UTC (304 KB)
[v2]
Tue, 13 Oct 2026 21:59:55 UTC (278 KB)
来源：arXiv cs.AI · arxiv.org