---
id: 20261018-T0-06
title: "把“继续推理还是收手”变成三选一：置信度校准的推理扩展控制策略"
title_en: "Confidence-Calibrated Inference Expansion Frames Reasoning as a Three-Action Control Problem"
url: https://ai.daily.yangsir.net/daily/20261018-T0-06
issue_date: 2026-10-18
publish_date: 2026-10-17T04:00:00.000Z
category: research
source_name: "arXiv cs.AI"
source_url: https://arxiv.org/abs/2610.20836
---

# 把“继续推理还是收手”变成三选一：置信度校准的推理扩展控制策略

论文把评估器引导的推理看作每一步的局部控制问题：信任当前答案、返回备选，还是投入更多推理算力。作者将其形式化为算力感知的三动作控制器，用预测值对终止动作打分，并用置信度校准来决定何时扩展推理。这样做的好处是把推理预算的分配从固定策略变成按步骤动态决策。对推理成本敏感的部署场景，这类控制器提供了在准确率与算力开销之间做显式权衡的框架。

## English Version

**Confidence-Calibrated Inference Expansion Frames Reasoning as a Three-Action Control Problem**

The paper frames evaluator-guided reasoning as a local control problem at every step: trust the current answer, return a fallback, or spend more inference. The authors formulate it as a compute-aware three-action controller that scores terminal actions by predicted values and uses confidence calibration to decide when to expand inference. The benefit is turning inference budget allocation from a fixed policy into a per-step dynamic decision. For cost-sensitive deployments, this kind of controller offers an explicit framework for trading accuracy against compute spend.

---

**来源**：[arXiv cs.AI](https://arxiv.org/abs/2610.20836)

**详情页**：https://ai.daily.yangsir.net/daily/20261018-T0-06

---

*智语观潮 · Daily — https://ai.daily.yangsir.net/llms.txt*