arXiv cs.AI· Weida Liang, Kenji Kawaguchi·· 2 天前AI 评分22
面向评估器引导测试时推理的置信度校准推理扩展
Confidence-Calibrated Inference Expansion for Evaluator-Guided Test-Time Reasoning
AI 导读
研究将评估器引导推理建模为计算感知的三动作控制器，在信任当前答案、回退与继续扩展之间做选择，并按预测答案成功率与基于 rollout 的延续成功代理打分，显式计入计算成本。在留出的宽度扩展与修订任务上，该控制器相比校准选定的固定策略提升了观测效用，同时将归一化在线计算量分别降低约 30% 和 16%。在四个 Qwen3-8B 随机种子上，两种设置的计算量均低于拟合 Q，效用差异较小且方向不一。
正文 · 原文
Abstract:Evaluator-guided reasoning presents a local control problem at every step: trust the current answer, return a fallback, or spend more inference. We formulate this choice as a compute-aware three-action controller, scoring terminal actions by predicted answer success and expansion by a rollout-based continuation-success surrogate, with compute cost priced explicitly. A local action-gap analysis characterizes when score errors can change a decision, while an expected regret bound accounts for errors along the realized expansion path. Guided by this analysis, we learn and calibrate action-specific confidence estimates, with local uncertainty penalties for ambiguous decisions. On held-out width expansion and revision tasks, the controller improves observed utility over calibration-selected fixed policies while reducing normalized online compute by about 30% and 16%, respectively. Across four Qwen3-8B seeds, compute is lower than fitted Q in both settings, while utility differences are small and mixed. The results frame evaluator-guided reasoning as resource allocation under uncertainty, rather than a uniformly stronger reasoner.
| Comments: | Accepted at NeurIPS 2026 |
| Subjects: | Artificial Intelligence (cs.AI); Computation and Language (cs.CL); Machine Learning (cs.LG) |
| Cite as: | arXiv:2610.20836 [cs.AI] |
| (or arXiv:2610.20836v1 [cs.AI] for this version) | |
| https://doi.org/10.48550/arXiv.2610.20836 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Weida Liang [view email]
[v1]
Wed, 14 Oct 2026 19:25:23 UTC (115 KB)
来源：arXiv cs.AI · arxiv.org