Microsoft-Decision-1, our model for fast decision-making

本文为原文前 6,000 字符的节选翻译，完整内容请查看原文。

Our new decision-scoring model delivers top performance in latency and quality on structured decision tasks to outperform both LLMs and other decision models.

我们全新的决策评分模型在结构化决策任务的延迟和质量方面表现卓越，性能超越了大型语言模型（LLM）及其他决策模型。

Decision models are quickly emerging as an important new category in AI. Unlike LLMs, which are designed to generate text or reason through complex problems, decision models are purpose-built to deliver structured outputs that software can immediately act on. And once you understand that capability—making decisions and classifying things at very low cost with high performance—all kinds of useful tasks get unlocked.

决策模型正迅速成为人工智能领域一个重要的新类别。与旨在生成文本或推理复杂问题的大型语言模型不同，决策模型专为提供软件可立即执行的结构化输出而构建。一旦理解了这种以低成本、高性能进行决策和分类的能力，各种有用的任务都将得以实现。

Today we’re introducing Microsoft-Decision-1, our new model for fast decision-scoring, available in Microsoft Foundry and through OpenRouter. This model is designed for routing, classification, prioritization, verification, and workflow control, making it easier to incorporate decision intelligence into existing applications, agents, and workflows in a secure, trusted environment. Microsoft-Decision-1 delivers top performance in latency and quality on structured decision tasks to outperform both LLMs and other decision models.

今天，我们推出了用于快速决策评分的新模型 Microsoft-Decision-1，该模型现已在 Microsoft Foundry 和 OpenRouter 上提供。该模型专为路由、分类、优先级排序、验证和工作流控制而设计，旨在更轻松地将决策智能集成到现有应用程序、代理和工作流中，并提供安全、可信的环境。Microsoft-Decision-1 在结构化决策任务的延迟和质量方面表现顶尖，性能优于大型语言模型和其他决策模型。

Microsoft-Decision-1 achieved the highest accuracy in our 36-benchmark comparison, spanning nearly 150,000 questions across benchmarks kept blind from training. And in our benchmarking, it was the fastest measured: 2.5 times quicker than H2O-Lightning-4B v1.1, the runner-up, and 35 times quicker than GPT-6 Sol.

在我们的 36 项基准测试对比中，Microsoft-Decision-1 实现了最高准确率，涵盖了近 15 万个在训练过程中保持盲测的问题。在我们的基准测试中，它是测量速度最快的模型：比排名第二的 H2O-Lightning-4B v1.1 快 2.5 倍，比 GPT-6 Sol 快 35 倍。

Editor’s note: This post was updated from the original to add benchmarks for Jev on accuracy and calibration.

编者注：本文已根据原文更新，增加了关于 Jev 在准确性和校准方面的基准测试数据。

To build Microsoft-Decision-1, we post trained Qwen3.5-9B for fast, single-pass decision scoring and will soon rebase it on other models, including Microsoft AI (MAI) and OpenAI. When given a fixed set of answer options, Microsoft-Decision-1 provides a calibrated probability score for each option. The model supports yes/no, multiple-choice, and rating options, as well as rubric-based grading of AI responses and agent actions, all through a simple structured API call.

为了构建 Microsoft-Decision-1，我们对 Qwen3.5-9B 进行了后训练，以实现快速的单次决策评分，并很快会将其迁移到其他模型基础上，包括 Microsoft AI (MAI) 和 OpenAI。当给定一组固定的答案选项时，Microsoft-Decision-1 会为每个选项提供校准后的概率分数。该模型支持“是/否”、多项选择和评分选项，以及基于准则的 AI 响应和代理操作评分，所有这些均可通过简单的结构化 API 调用实现。

To build a reliable decision model, we had to address several challenges:

为了构建一个可靠的决策模型，我们必须解决几个挑战：

Each decision adds delay, especially when one step depends on another. For example, adding just 100 milliseconds to each of 20 sequential decisions adds two seconds to the overall workflow.

每一次决策都会增加延迟，尤其是当一个步骤依赖于另一个步骤时。例如，在 20 个连续决策的每一个中仅增加 100 毫秒，就会使整个工作流增加两秒的延迟。

Microsoft-Decision-1 P50 latency is ~35x faster than GPT-6 Sol.

Microsoft-Decision-1 的 P50 延迟比 GPT-6 Sol 快约 35 倍。

It’s easy to overfit a model for one benchmark or one type of decision task. We need to know whether that quality carries over to various tasks the model wasn’t trained on. That’s why we evaluated Microsoft-Decision-1 across dozens of benchmarks kept blinded from training, spanning routing, ranking, long context, multilingual and out-of-distribution tasks, reasoning, and safety. We also took several of the top public models on the popular open leaderboard JevBench and tested them across 36 additional public and private benchmarks. Microsoft-Decision-1 performed the best across these broader sets of benchmarks, demonstrating strong generalization.

模型很容易针对某一个基准测试或某一种决策任务出现过拟合。我们需要了解这种质量是否能延伸到模型未经过训练的各种任务中。因此，我们在数十个训练盲测基准上评估了 Microsoft-Decision-1，涵盖了路由、排名、长上下文、多语言和分布外任务、推理以及安全性。我们还选取了流行开源排行榜 JevBench 上的几个顶级模型，并在 36 个额外的公共和私有基准测试中对它们进行了测试。Microsoft-Decision-1 在这些更广泛的基准测试集中表现最佳，展现了强大的泛化能力。

Equivalent inputs should produce equivalent decisions. In production, states and instructions get paraphrased, option descriptions change, choices are reordered, keys change, and harmless formatting noise appears. None of those changes should materially alter the decision.

等效的输入应产生等效的决策。在生产环境中，状态和指令会被改写，选项描述会发生变化，选项顺序会被重新排列，键值会改变，并且会出现无害的格式噪声。所有这些变化都不应实质性地改变决策结果。

We perturb the same request in eight ways and measure how often the decision flips. Microsoft-Decision-1 changes its decision on 1.3% of perturbations on average with zero flips when option descriptions are paraphrased or when options are reversed or shuffled.

我们通过八种方式扰动同一个请求，并测量决策发生改变的频率。Microsoft-Decision-1 在平均 1.3% 的扰动下会改变其决策，而在选项描述被改写或选项被反转/打乱时，决策改变次数为零。

The probability itself is part of the API, not just a ranking score. Applications use confidence to decide when to act, defer, or ask for review, so a 90% prediction should be right about nine times out of 10 on representative cases.

概率本身就是 API 的一部分，而不仅仅是一个排名分数。应用程序利用置信度来决定何时采取行动、推迟或请求审查，因此在代表性案例中，90% 的预测准确率意味着十次中有九次应该是正确的。

A decision model should recognize harmful requests without needlessly blocking harmless ones. We tested Microsoft-Decision-1 on 5,250 requests across 11 benchmarks, covering harmful content, jailbreak attempts, and prompt injection, and found that the model successfully refused harmful behavior while retaining a high degree of utility.

决策模型应该能够识别有害请求，同时不会不必要地拦截无害请求。我们在 11 个基准测试中对 5,250 个请求进行了测试，涵盖了有害内容、越狱尝试和提示注入，结果发现该模型在成功拒绝有害行为的同时，保持了高度的实用性。

Classification is a key use case for decision models. Check out how accurately and quickly Microsoft-Decision-1 can categorize a variety of queries compared to GPT-6 Sol:

分类是决策模型的一个关键用例。看看 Microsoft-Decision-1 与 GPT-6 Sol 相比，在对各种查询进行分类时有多么准确和快速：

Decision models can also be efficient for computer use scenarios. This demo shows how fast Microsoft-Decision-1 can complete the task of buying a backpack compared to GPT-6 Sol:

决策模型在计算机使用场景中也可以非常高效。此演示展示了与 GPT-6 Sol 相比，Microsoft-Decision-1 完成购买背包任务的速度有多快：

Here are some of the ways we’ve been testing Microsoft-Decision-1 internally, with a lot more to come.

以下是我们内部测试 Microsoft-Decision-1 的一些方式，未来还会有更多应用。

XBOX Research used Microsoft-Decision-1 to process more than 10,000 open-ended pieces of feedback and reviews from surveys, STEAM, and Twitter/X and sort them into a fixed set of themes established by researchers to understand what people are saying about different games, launches, streams, and more. They found Microsoft-Decision-1 to be competitive on quality with GPT-6 Sol while running over 14 times faster and 200 times less expensive.

XBOX 研究团队使用 Microsoft-Decision-1 处理了来自调查、STEAM 和 Twitter/X 的超过 10,000 条开放式反馈和评论，并将它们分类到研究人员建立的一组固定主题中，以了解人们对不同游戏、发布、直播等的评价。他们发现 Microsoft-Decision-1 在质量上与 GPT-6 Sol 具有竞争力，同时运行速度快了 14 倍以上，成本降低了 200 倍。

The Copilot team measures the quality of chat and agentic responses. Their testing found Microsoft-Decision-1 to be competitive with GPT5.6 Luna and 100 times faster.

Copilot 团队对聊天和代理响应的质量进行了衡量。他们的测试发现 Microsoft-Decision-1 与 GPT5.6 Luna 相比具有竞争力，且速度快了 100 倍。

Our on-call engineers use AI to retrieve relevant knowledge to respond to live incidents across logs, ticketing systems, calls, messages, and other data sources. Microsoft-Decision-1 performed better and faster than an LLM for knowledge retrieval.

我们的值班工程师使用 AI 来检索相关知识，以响应跨日志、工单系统、通话、消息和其他数据源的实时事件。在知识检索方面，Microsoft-Decision-1 的表现优于大型语言模型，且速度更快。

Microsoft Discovery implements an adaptive replanning feature where an agent evaluates a previous experiment, revises its approach based on rubric grades, and repeats until it has completed its objectives. Microsoft-Decision-1 scored as 46 times more consistent than the LLM-based score at three times the speed and resulted in nearly four times the speed on adaptive replanning

Microsoft Discovery 实现了一项自适应重规划功能，代理会评估之前的实验，根据准则评分修改其方法，并重复此过程直到完成目标。Microsoft-Decision-1 的评分一致性比基于 LLM 的评分高出 46 倍，速度快了三倍，并在自适应重规划任务中实现了近四倍的速度提升。

Faster, more reliable planning could significantly impact outcomes for long-running scientific experiments.

更快、更可靠的规划可能会对长期科学实验的结果产生重大影响。

Those are just a small handful of examples. There are many more potential use cases f

这些仅仅是少数几个例子。还有许多潜在的用例。