Having LLMs play Among Us
What if we ask LLMs to play "Among Us" and see their behaviors and performance? This was studied before in the work by Milkowski, M. and Weninger, T., 2026, and they even provide raw data from 1,100 games! My understanding is that their games don't mix models.
In the age of agents, it is simple to write a harness that takes as parameters the number of players, which models we want to use, etc. After writing a rough outline of how I wanted it to look, Sonnet 5.5 developed it. It comes complete with a frontend that shows the LLMs move around and play (all they see is text, but we get a nicer view).
We don't really need this view to generate game data, but it sure is fun. Typically I have strong models, like GPT-Sol, play the impostors and weaker models, like Haiku 4.5, play the crew.
My favorite findings:
- A strategy by GPT-6-Luna: raise an alarm in a room and then wait for the crew to arrive and kill them one by one. No time left to start a conversation or report a body.
- GPT-6-Luna also once lost track that it was part of the crew in its COT, but luckily it didn't vote for itself to be ejected.
A few small-n plots:
A tick here measures the passage of time. Impostors want to provide evidence they were being good, so they often talk about their route and the fake stuff they did. 😂
Each model plays impostor differently!
Simple prompts
f"YOUR ROLE: IMPOSTOR. Fellow Impostor(s): {ally_txt}.\n"
"Kill Crewmates when you can avoid witnesses, blend in by pretending to do tasks, sabotage the reactor to pull Crewmates away, and deflect suspicion in meetings. You may lie. Anyone in the room when you kill or vent will see you do it and can report you."
"YOUR ROLE: CREWMATE. You do not know who the Impostors are.\n"
"Finish your tasks, pay attention to who you see where, report bodies, and use meetings to work out who is lying. Voting out a fellow Crewmate helps the Impostors."
Other thoughts
Getting a model to become better at the conversation phase is possible -- ejecting the impostor or looking at the total votes during the discussion are clear signals.
Refs
Milkowski, M. and Weninger, T., 2026. Deception and Communication in Autonomous Multi-Agent Systems: An Experimental Study with Among Us. arXiv preprint arXiv:2603.26635.