Training Among Us Bots: The Crew That Stopped Fixing the Reactor
I train Among Us bots by self-play: the crew and the impostors are both learners, and they play tens of thousands of games against each other and against saved copies of their earlier selves. I run the training in leagues, each one a roughly ten-hour run that starts from the bots the previous league ended with. The fifth league looked healthy for ten thousand games and then fell over. Between…
Among Us bots were trained using self-play, with both crewmates and impostors learning through tens of thousands of games against each other and previous versions of themselves. The fifth league of training initially performed well before abruptly falling apart, with the impostors' win rate increasing dramatically and the crew's task completion rate dropping significantly. The impostors began preferring to sabotage the reactor instead of killing crewmates, which improved their win rate.
Researchers created a probe to analyze the crew's decision-making during sabotage situations. They found that during the fifth league, crewmates placed significantly less importance on getting closer to the fix compared to their previous league. The crew's reward system – based on game results, task completion percentage, being alive at the end, and task steps completed by the crew – did not directly incentivize fixing the reactor.
As a result, the crew started placing less value on fixing the reactor, leading to more meltdowns and losses. The impostors' strategy of sabotaging the reactor was more directly rewarding, as it directly contributed to their victory.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
