The Virtual Roundtable: Multi-Agent Personas Simulating the Dynamics of Human Brainstorming
This paper presents a multi-agent architecture that simulates human roundtable brainstorming through diverse AI personas and an agentic facilitator to overcome groupthink, effectively generating and evolving high-quality, diverse ideas through structured divergent and convergent thinking phases.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine you are trying to solve a tricky problem, like designing a new pair of smart glasses. Usually, you'd gather a group of real people in a room to brainstorm. But real people have problems: they might be too shy to speak up, they might just agree with the loudest person (groupthink), or they might run out of time before they get to the good ideas.
This paper introduces a virtual brainstorming room filled with AI characters (called "personas") instead of real humans. Think of it like a high-tech, digital version of a roundtable discussion where every participant is a distinct, carefully crafted character with their own unique personality, job, and way of seeing the world.
Here is how the system works, broken down into simple parts:
1. The Cast of Characters
The system creates 13 different AI "actors." One might be a "Busy Working Parent" who cares about efficiency and time. Another is a "Tech Enthusiast" who loves battery life and cloud computing. A third is an "Accessibility Advocate" who worries about whether the glasses work for blind people.
- The Analogy: Imagine a movie set where the director (the system) casts a diverse group of actors. Each actor stays in character the whole time, bringing their specific life experience to the conversation.
2. The Three-Act Play
The brainstorming session isn't just a free-for-all chat; it follows a strict script with three acts to ensure creativity doesn't get killed by early criticism.
- Act 1: The Warm-Up (Discussion): The characters talk about the problem to understand each other. No ideas are written down yet. It's like the actors improvising to get into the scene. They learn what the "Travel Enthusiast" needs versus what the "Homebody" needs.
- Act 2: The Explosion (Ideation): Now, the characters shout out ideas. The rule is: No bad ideas allowed. You can't criticize anyone yet. The goal is to get as many ideas as possible on the whiteboard. This is like a "divergent" phase where the group spreads out to find every possible angle.
- Act 3: The Vote (Evaluation): Once the ideas are on the board, the characters switch gears. Now they act as judges. They vote on which ideas are best based on their own values. This is the "convergent" phase where the group narrows down to the winners.
3. The Invisible Moderator
There is a special AI called the Facilitator. It doesn't offer any ideas of its own. Instead, it acts like a strict but helpful teacher or a referee.
- What it does: It watches the room. If one character is talking too much, the Facilitator gently nudges them to let others speak. If the group starts agreeing too quickly (echo chamber), the Facilitator asks a tough question to shake things up. It ensures everyone gets a fair turn to speak.
4. The "Thought Bubble" Secret
One of the coolest features is that every character has a private thought bubble.
- How it works: When someone says something out loud, every other character first thinks about it privately in their own head, then says something out loud.
- The Result: This mimics how humans actually think. We hear a comment, process it internally, and then respond. The system tracks this "lineage," showing exactly how one idea grew from a comment made by someone else, creating a family tree of ideas.
5. What They Found (The Results)
The researchers tested this system by asking the AI characters to design smart glasses. They ran many sessions, changing only one thing: how long the "Warm-Up" (discussion) phase lasted.
- Longer Discussions = Deeper Ideas: When the characters talked for a longer time before generating ideas, the final ideas were much more sophisticated. The ideas were "cross-pollinated," meaning they combined insights from many different characters. For example, a "Tech" idea about battery life got mixed with a "Parent" idea about safety, creating a better, more complete product.
- Longer Discussions ≠ Newer Ideas: However, making the discussion longer didn't make the ideas more different from each other. The variety of ideas was determined by who was in the room (the mix of characters), not how long they talked. If you only have tech people, you get tech ideas, no matter how long they talk. If you add a parent and a traveler, you get a wider variety of ideas.
The Big Takeaway
This system proves that you can simulate a high-quality human brainstorming session using AI.
- The Secret Sauce: It's not just about having AI chat; it's about having distinct characters who talk to each other, listen to each other, and are guided by a moderator to keep the conversation productive.
- The Lesson: If you want deep, well-thought-out ideas, let your team (or your AI agents) talk and debate for a while before they start trying to solve the problem. But if you want a wide variety of wild ideas, you need to make sure you have the right mix of different people (or personas) in the room to begin with.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.