AWS 提出 AEM：多轮对话智能体的逐轮可分解评估指标

AI 摘要

AWS Machine Learning Blog 发布文章介绍 Agent Evaluation Metric（AEM），一种面向多轮对话的、可分解的逐轮评估方法。文章指出多轮智能体中的早期错误会级联污染后续轮次，而任务级或整体评分无法定位根因。AEM 首先实例化正确性维度，将其拆分为真实性和完整性两个子指标，并支持逐轮测量与跨轨迹组合，未来可扩展到安全、指令保持等维度。

正文节选

Agent Evaluation Metric for multi-turn conversations Multi-turn agents fail in ways that single-turn evaluation misses: one early mistake quietly corrupts every later turn. This post introduces the Agent Evaluation Metric (AEM), a decomposable, turn-level way to measure agent quality. We apply it to its first dimension, correctness. We show how AEM pinpoints the one turn that caused a failure and separates it from the turns that merely inherited the problem. The correctness challenge in multi-tu

发布时间：2026-09-10 23:55

抓取时间：2026-09-11 00:17

来源机构：AWS