跳转至

文章背景与核心概要

在将任务委派给线上会议中的大语言模型（LLM）时，智能代理（Agents）往往由于缺乏情境感知（特别是无法识别何时该发言）而频繁失效。如果没有一个结构化的机制来追踪对话立场、话题覆盖范围以及当前谁拥有发言权，仅靠提示词（Prompt-only）的代理会错过超过一半的发言机会。

为了解决这一问题，本文作者推出了 CAPA（Collaborative Agent Predictive Architecture，协同代理预测架构），这是一个专为线上会议委派设计的新颖框架。CAPA 将静默率从 51.4% 大幅降低至 2.5%，使有效发言召回率（credited recovery）翻倍，同时保持了仅 0.6% 的极低幻觉率。

Speak for Me: Giving LLMs the Situational Awareness to Participate in a Meeting

arXiv ID: arXiv:2609.03923 [cs.AI]
Accepted at: EMNLP 2026 Main
Authors: Muneeb Khan, Frederic Kirstein, Terry Ruas, Bela Gipp
Submission Date: September 3, 2026

📌 执行摘要

When delegating tasks to Large Language Models (LLMs) in online meetings, agents frequently fail due to a lack of situational awareness—specifically, recognizing when to speak. Without a structured mechanism to track conversational stances, topic coverage, and who holds the floor, prompt-only delegates miss over half of all talking opportunities.

To solve this, the authors introduce CAPA (Collaborative Agent Predictive Architecture), a novel framework designed for online meeting delegation. CAPA dramatically reduces silence rates from 51.4% to 2.5%, doubles credited recovery, and maintains an exceptionally low hallucination rate of 0.6%.

🔍 核心架构：CAPA（协同代理预测架构）

CAPA 利用模块化流水线实时处理会议动态：

CAPA utilizes a modular pipeline to process meeting dynamics in real time:

  1. 感知器（Perceiver）： 根据观察到的每一个发言轮次更新当前会议状态。
    1. Perceiver: Updates the ongoing meeting state from each observed turn.
  2. 预测器（Predictor）： 预测对话未来的演变方向。
    1. Predictor: Forecasts how the conversation is likely to unfold.
  3. 控制器（Controller）： 决定代理是否应当发言，并挑选出合适的命题呈现。
    1. Controller: Decides whether the agent should speak and selects the appropriate proposition to surface.
  4. 生成器（Generator）： 按照被委派参与者独特的个人风格来组织所选贡献的表述。
    1. Generator: Phrases the chosen contribution matching the delegated participant's unique personal style.
  5. 裁判与校准器（Judges & Recalibrator）： 两个裁判针对下一个实际观察到的发言轮次对预测和行动进行评分，校准器则更新内部会议状态以供后续决策使用。
    1. Judges & Recalibrator: Two judges score the forecast and action against the next actual observed turn, and a Recalibrator updates the internal meeting state for future decisions.

📊 评估与结果

  • 评估方案： 作者引入了一种片段级（Episode-level）的评估方案，用于打分代理是否、何时以及围绕实际观点单元贡献了什么。模式受限的 LLM 裁判与人工标注高度一致，科恩 kappa 系数（Cohen's kappa）达到 0.71。
    • Evaluation Protocol: The authors introduce an episode-level protocol scoring whether, when, and what a delegate contributes around actual idea units. Schema-constrained LLM judges align closely with human annotations at Cohen's kappa = 0.71.
  • 在 AMI 语料库（137场会议）上的表现：
    • Performance on AMI Corpus (137 Meetings):
  • 静默率： 从 51.4% 大幅降至 2.5%。 > * Silence Rate: Reduced from 51.4% down to 2.5%.
  • 有效发言召回率： 从 26.1 翻倍至 52.2。 > * Credited Recovery: Doubled from 26.1 to 52.2.
  • 幻觉率： 保持在 0.6% 的极低水平。 > * Hallucination Rate: Kept minimal at 0.6%.
  • 消融实验： 机制消融实验证实，维护一个显式的会议状态是弥补识别差距的主要杠杆——这是单靠原始上下文扩展（Raw-context scaling）无法实现的功能。
    • Ablations: Mechanism ablations confirm that maintaining an explicit meeting state is the primary lever closing the recognition gap—a feat raw-context scaling alone cannot achieve.

🔗 链接与资源