This note is essentially an extension and running log of research and ideas inspired by my Questions about LLMs in Groupchats post
Multi-Party Conversational Agents: A Survey: https://arxiv.org/pdf/2505.18845v1
3 main concerns in regards to MPCAs
- Theory of mind / State of Mind Modeling
- Semantic understanding of the content of the conversation
- Predict future conversation flow — Agent Action Modeling
For thinking about State of Mind Modeling they reference the OCEAN model for personality traits and mention 4 different components
- emotion recognition
- Is something positive or negative
- engagement detection
- How often is someone talking
- personality recognition
- identify personality traits of participants from conversation
- Generally using Big-Five Personality Model
- conversational intentions
- Classify the intent of a message (question vs statement)
Semantic Understanding is also is split into 4 tasks that relate to an agent’s ability to understand the meaning and flow of a conversation
- Conversation disentanglement
- Mainly related to if multiple people are talking over each other can you split it up into the speakers.
- Less relevant in chat settings I think depending on the setup
- Dialogue Summarization
- Discourse Structure Analysis
- Representation Learning
Agent Action modeling is split into 3 tasks that all kind of revolve over who speaks next.
- Turn Detection
- Addresse Selection
- Agent Response
Stream of consciousness Notes
There something that feels very skeumorphic about how all of these different papers or related ideas try to map a human conversation to agents. Mainly in regards to time.
I don’t really care about conversational disentanglement too much with text based chat models. That’s not really a concern in things like discord or text messages. Thinking more how can you emulate a conversation that looks like a chat application.
Time feels like the key here. After someone has said something in a groupchat oftentime multiple people will start typing or thinking of responding. With LLMs this is problematic because if you just let everyone try to start talking immediately the smallest model (weightwise) is going to just be faster and more capable of drowning out everyone.
So what’s a way to emulate a pause after a message? Is there a way to have some kind of tick system similar to a video game?
A rough thought here is after each message is sent, each agent - individually - processes the message, thinks about if they want to respond, indicates if they want to respond, and then generates its response. Then maybe there is reconciliation phases where each agent sees what other agent wants to talk (emulating the seeing a …typing indicator). Each agent can then decide if they want one of the people typing to talk instead of them. There are several rounds of reconciliation or maybe just one. After a round of reconciliation whoever is remaining gets to send their message with a randomness in the order or maybe some kind of weight based on the length of the message (should take longer to type a longer message so the shorter message should be sent first).
How would this look in something like werewolf or mafia? Could that be a good testbed? Instead of a timer for discussion it’s a number of ticks