主动式服务智能体：统一的决策框架、方法与评估

文章背景与核心概要

尽管现代大语言模型（LLM）智能体在规划、工具调用和修改外部状态方面表现出色，但它们传统上依赖明确、被动的用户指令作为固定的起点。本文引入了主动式服务智能体（Proactive Service Agents），这是一种将决策过程向前推进的范式转变。

主动式智能体不再被动等待提示词，而是从不完整的环境和用户信号中推断服务机会，自主决定是保持沉默、询问、协助还是采取行动，并在干预成本、误解风险、越权代价和隐私保护之间进行仔细权衡。本文构建了统一的决策框架，梳理了相关方法论，并为这一前沿方向奠定了坚实的评估基础。

执行摘要 / Executive Summary

While contemporary Large Language Model (LLM) agents excel at planning, tool invocation, and modifying external states, they traditionally rely on explicit, reactive user instructions as a fixed starting point. This paper introduces Proactive Service Agents, a paradigm shift that moves decision-making upstream.

尽管现代大语言模型（LLM）智能体在规划、工具调用和修改外部状态方面表现出色，但它们传统上依赖明确、被动用户指令作为固定的起点。本文引入了主动式服务智能体，这是一种将决策过程向前推进的范式转变。

Instead of waiting for prompts, proactive agents infer service opportunities from incomplete environmental and user signals, decide whether to remain silent, ask, assist, or act, and carefully balance interruption, misunderstanding, overreach, and privacy costs.

主动式智能体不再等待提示词，而是从不完整的环境和用户信号中推断服务机会，自主决定是保持沉默、询问、协助还是采取行动，并在仔细权衡打扰成本、误解风险、越权代价和隐私隐私开销。

框架核心亮点 / Key Highlights of the Framework

• Operational Definition: Centers heavily on initiative, formulating the proactive service problem as a partially observable sequential decision process constrained by authorization and risk.

  • 操作性定义： 重点聚焦于“主动性”，将主动式服务问题建模为一个受授权和风险约束的部分可观测序列决策过程（partially observable sequential decision process）。

• Unified Structured Actions: Represents timing, content, and delivery within a single structured action format while factoring in the option value of waiting, the decision value of clarifying questions, and feedback-induced state changes.

  • 统一的结构化动作： 在单一的结构化动作格式中表示时机、内容和交付方式，同时兼顾等待的期权价值（option value）、澄清问题的决策价值以及反馈引起的状变。

• The Decision Pipeline: Organizes existing proactive methodologies into four sequential stages:

• State and need estimation
• Intervention gating
• Action construction
• Feedback adaptation

  • 决策流水线： 将现有的主动式方法论组织为四个连续的阶段：

• 状态与需求估计
• 干预门控（Intervention gating）
• 动作构建

• 反馈自适应

• Policy-Construction Mechanisms: Outlines prescribed, predictive, model-based, and return-optimizing mechanisms as nonexclusive building blocks.

  • 策略构建机制： 概述了规定式（prescribed）、预测式、基于模型以及优化回报的机制，将其作为非排他性的构建模块。

• Diverse Domain Normalization: Normalizes decision units and three-axis evidence descriptors across multiple resource types, including streaming dialogue, screen navigation, video, software engineering, and human-agent collaboration.

  • 多领域归一化： 跨多种资源类型（包括流式对话、屏幕导航、视频、软件工程和人机协作）对决策单元和三轴证据描述符进行归一化。

• Evaluation Metrics: Formalizes robust metrics for triggering, timing, calibration, user burden, safety, and policy value.

  • 评估指标： 针对触发、时机、校准、用户负担、安全性和策略价值，形式化了一套鲁棒的评估指标。

结论与见解 / Conclusion & Insights

The paper demonstrates why traditional offline classification performance alone is insufficient for predicting real-world deployment benefits, and proves that long-term memory is not a defining prerequisite for proactivity.

本文论证了为什么单纯的离线分类性能不足以预测现实世界的部署收益，并证明了长期记忆并不是实现主动性的决定性前提。

Ultimately, reliable proactive service requires: * Calibrated incremental intervention value * Verifiable authorization mechanisms * Recoverable execution strategies * Grounded counterfactual evidence

归根结底，可靠的主动式服务需要： * 经校准的增量干预价值 * 可验证的授权机制 * 可恢复的执行策略 * 有依据的反事实证据