ThinkJev: One Thought, Every Budget for Calibrated Typed Decisions

查看原文 / View Original 编辑 / Edit

Computer Science > Artificial Intelligence arXiv:2610.37732 (cs) [Submitted on 27 Oct 2026] Title:ThinkJev: One Thought, Every Budget for Calibrated Typed Decisions Authors:Xiaolu Ding, Kai Yang, Dongdong Liu View a PDF of the paper titled ThinkJev: One Thought, Every Budget for Calibrated Typed Decisions, by Xiaolu Ding and 2 other authors View PDF HTML (experimental)

计算机科学 > 人工智能 arXiv:2610.37732 (cs) [提交于 2026 年 10 月 27 日] 标题：ThinkJev: One Thought, Every Budget for Calibrated Typed Decisions（ThinkJev：用于校准类型化决策的单一思维与全预算方案）作者：Xiaolu Ding, Kai Yang, Dongdong Liu 查看论文 PDF，标题为 ThinkJev: One Thought, Every Budget for Calibrated Typed Decisions，作者为 Xiaolu Ding 及其他 2 位作者查看 PDF HTML（实验性）

Abstract:Many uses of language models end in a typed decision: a choice among a closed set of options, whose probabilities code can act on. Reading those probabilities from one forward pass is fast but often wrong on decisions that take several steps, such as arithmetic, counting, and date comparison. Letting the model think first fixes most of these errors, but costs seconds per decision instead of milliseconds. To pay for thinking only where it helps, ThinkJev builds a fast and a slow system from one open hybrid-thinking model that can answer directly or after a thought.

摘要：语言模型的许多用途最终都会产生类型化决策：即在封闭的选项集中进行选择，代码可以根据这些选项的概率采取行动。通过一次前向传递读取这些概率速度很快，但在涉及算术、计数和日期比较等需要多个步骤的决策时，往往会出错。让模型先进行思考可以修复大多数此类错误，但每次决策需要耗费数秒而非毫秒。为了仅在有帮助的情况下才投入思考成本，ThinkJev 利用一个开放的混合思维模型构建了一个快速系统和一个慢速系统，该模型既可以直接回答，也可以在思考后回答。

System 1 reads the model’s probabilities for the option letters when it answers directly; System 2 reads the same letters after the model has thought on the same cached prompt, so escalating repeats no work and both systems give probabilities over the same options. Because the letters can be read after any number of thought tokens, one sampled thought also gives the answer under every thinking budget. Without training, ThinkJev lets System 1 answer alone only where a statistical test on labeled dev data guarantees that its accepted answers stay under a target error rate.

系统 1 在直接回答时读取模型对选项字母的概率；系统 2 在模型对相同的缓存提示进行思考后读取相同的字母，因此升级过程不会重复工作，且两个系统都能针对相同的选项给出概率。由于可以在任意数量的思维标记（thought tokens）之后读取字母，因此一次采样的思维也能在各种思维预算下提供答案。在无需训练的情况下，ThinkJev 仅在标记开发数据上的统计测试保证其接受的答案保持在目标错误率以下时，才允许系统 1 单独回答。

At a 5% target, Qwen3.5-27B answers 24.5% of 8359 test decisions this way with a realized error of 3.9% (Qwen3.5-4B and 9B: 6.5% and 0.0%). For the rest, stopping each thought as soon as its answer is confident reaches 81.6% accuracy, against 82.7% for always thinking, with 72% of the thought tokens. Choosing when to think from System 1’s confidence alone saves less, because the decisions System 1 is least sure of are often not the ones thinking fixes.

在 5% 的目标下，Qwen3.5-27B 以这种方式回答了 8359 个测试决策中的 24.5%，实际错误率为 3.9%（Qwen3.5-4B 和 9B 分别为 6.5% 和 0.0%）。对于其余部分，一旦答案确定就停止思考，准确率可达 81.6%，而始终进行思考的准确率为 82.7%，且仅使用了 72% 的思维标记。仅根据系统 1 的置信度来选择何时思考节省的成本较少，因为系统 1 最不确定的决策往往不是通过思考就能修复的决策。

On the public items of JevBench, an independent benchmark of typed-decision services, Qwen3.5-27B thinks on 29.4% and is not distinguishable in accuracy from the two thinking APIs listed (exploratory). Code and data generators are available at this https URL.

在类型化决策服务的独立基准测试 JevBench 的公共项目上，Qwen3.5-27B 在 29.4% 的情况下进行思考，其准确率与所列出的两个思维 API（探索性）没有显著差异。代码和数据生成器可在此 URL 获取。