Original title: Speak for Me: Giving LLMs the Situational Awareness to Participate in a Meeting
Muneeb Khan, Frederic Kirstein, Terry Ruas, Bela Gipp · September 2026
If you send an AI bot to attend an online meeting for you, it usually struggles to know when to speak up. Standard prompt-based bots stay silent during more than half of the moments their human should have talked. Researchers built a system called CAPA to fix this. It listens to each turn of speech, tracks the room's current state, and predicts where the conversation is going. Then, it decides if it should talk, picks the right point to make, and words it in the absent human's personal style. Afterward, internal judges grade its choices and adjust its notes for future turns. Tested on real meetings, CAPA cut awkward silence from 51.4% down to 2.5% and doubled the amount of useful human ideas it recovered, while keeping made-up facts at just 0.6%. Tracking the structured meeting state made the difference, not just feeding the bot more conversation text.
Imagine you're in a group chat where someone is absent, and you're supposed to speak for them. It's hard to know when to jump in because you can't always tell if your input is needed or how the conversation will go. In this paper, the AI agent faces this exact challenge in online meetings. CAPA tracks the chat, predicts what others will say, and decides when to contribute, learning from each interaction to get better at representing the person.
In meetings, missing the right moment to speak wastes time and loses ideas. This matters for remote teams and anyone using AI assistants, making collaboration more efficient.
The abstract doesn't say if CAPA works for meetings outside the AMI corpus or in real-time settings. It also doesn't address scalability or cost. The failure mode shifts from missing chances to picking wrong topics, so it's not foolproof.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Speak for Me: Giving LLMs the Situational Awareness to Participate in a Meeting — Muneeb Khan, Frederic Kirstein, Terry Ruas et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.