JevMan
A game for exploring how AI models make decisions.
JevMan is a Pac-Man benchmark for AI decision models, built with Felix Wunderlich. The game gives models a familiar problem: navigate a maze, collect pellets, and avoid the ghosts. At each junction, a model chooses a direction from the current game state.
Watching the game makes the decisions easier to inspect. The board shows the model’s odds for each direction, and an activity log records its choices, latency, and what happened next. Repeated games feed into a leaderboard so the comparison covers more than one good run.
Visitors can watch models play or steer Pac-Man against AI-controlled ghosts. The project is open source, with a way to run games and submit results from other models.