Speak for Me: Giving LLMs the Situational Awareness to Participate in a Meeting
CAPA (Collaborative Agent Predictive Architecture) is introduced for online meeting delegation. On 137 AMI meetings, CAPA reduces the silence rate from 51.4% to 2.5% and doubles credited contributions. The evaluation protocol's schema-constrained LLM judges align with human annotations at Cohen's kappa = 0.71.
Researchers present CAPA, an architecture for online meeting delegation that gives LLMs situational awareness to decide when and what to contribute. CAPA uses a Perceiver, Predictor, Controller, Generator, and Recalibrator to track meeting state, forecast conversation, decide speaking actions, and phrase contributions. An episode-level evaluation protocol scores whether, when, and what a delegate contributes. On the AMI corpus, CAPA reduces silence rate from 51.4% to 2.5% and doubles credited contributions, with judge-human agreement at Cohen's kappa = 0.71.
CAPA combines a Perceiver for state updates, a Predictor for conversation forecasting, a Controller for speaking decisions, a Generator for style-matched phrasing, and a Recalibrator for feedback-driven state updates. The evaluation uses schema-constrained LLM judges with Cohen's kappa = 0.71 against human annotations, indicating reliable automated assessment of contribution quality.
This research addresses a key limitation in meeting AI agents: knowing when to speak. Reducing silence rate from 51.4% to 2.5% suggests significant improvement in agent participation, which could enhance virtual meeting assistants and delegation tools.
Improved meeting delegation agents could increase productivity for remote and hybrid work by ensuring absent participants' viewpoints are represented, potentially reducing meeting follow-ups and miscommunication.
Next signals include adoption of CAPA-like architectures in commercial meeting assistants, further validation on other meeting corpora, and improvements in real-time latency and personalization of speaking style.