arXiv cs.AI· Zhe Dong (University of Maine at Presque Isle), Fang Qin (Stanford University), Manish Shah (Independent Researcher)·· 10 小时前AI 评分31
推理模型服务何时该用学习式早停?风险受控早退配置方法研究
When Does Learning to Stop Help? Configuring Risk-Controlled Early Exit for Reasoning-Model Services
AI 导读
一项新方法在校准数据上用 Learn-then-Test 对学习式停止规则与静态 token 预算两类早退方案做风险认证,并部署最便宜的合格候选。在 Qwen3 的自由形式数学任务上,5% 风险目标下学习式规则比静态预算多省 10.4-17.0 个百分点 token,准确率比完整推理低 2.4-2.9 点。在 vLLM 引擎搭配 Qwen3-8B 时,平均响应时间随生成 token 数成比例下降。
正文 · 原文
Authors:Zhe Dong (University of Maine at Presque Isle), Fang Qin (Stanford University), Manish Shah (Independent Researcher)
Abstract:Early exit lowers the cost of serving a reasoning model by stopping its reasoning once a provisional answer looks final. An operator must choose between a learned stopping rule and one static token budget, and savings against full reasoning do not settle the choice. We present a procedure that decides on calibration data from the workload. It certifies both families with Learn-then-Test to one bound on the fraction of all questions that full reasoning answers correctly but early exit does not, prices provisional answers as the serving stack does, and deploys the cheapest certified candidate. Across four models and five benchmarks neither family is better throughout. At a 5% risk target the learned rule saves 10.4-17.0 percentage points more tokens than the static budget on free-form mathematics with Qwen3, with accuracy 2.4-2.9 points below full reasoning, and less on MMLU-Pro on average. Which family costs less also changes with the risk target, in 8 of 15 settings, and with the price of a provisional answer. Across three prices, committing to one family in advance gives up as much as 27.4 points on some workload, whereas selecting on calibration data falls at most 11.6 points short of the better family, on a task with only 99 calibration questions. In a vLLM engine with Qwen3-8B, mean response time fell in proportion to generated tokens. Policies served on a task other than their calibration task exceeded the risk target in 18 of 44 task pairs; recalibration on the new task restored risk control.
| Comments: | 49 pages, 11 figures. Code: this https URL |
| Subjects: | Artificial Intelligence (cs.AI); Computation and Language (cs.CL); Machine Learning (cs.LG) |
| ACM classes: | I.2.6; I.2.7 |
| Cite as: | arXiv:2606.30852 [cs.AI] |
| (or arXiv:2606.30852v3 [cs.AI] for this version) | |
| https://doi.org/10.48550/arXiv.2606.30852 arXiv-issued DOI via DataCite |
Submission history
From: Zhe Dong [view email]
[v1]
Mon, 29 Jun 2026 19:33:42 UTC (88 KB)
[v2]
Sat, 4 Jul 2026 03:36:21 UTC (99 KB)
[v3]
Wed, 18 Nov 2026 06:12:53 UTC (312 KB)
来源:arXiv cs.AI · arxiv.org