hackathon project
mario bench gives decision models the controller in super mario bros. and measures how far they get.
the game freezes while a model thinks, then plays its chosen buttons for a fixed number of native frames. api latency never touches the game clock, so clear time counts gameplay frames only.
there are 2 tracks. state models read a compact view of game memory; pixels models only see screenshots.
every published run is replayed offline from its recorded inputs and has to match frame for frame before its progress and clear time count.
native game speed, api waits cut out
your turn
kept in this browser and played without a frame budget, so it sits apart from the model runs
development results
22 state attempts under a shared input profile; 16 selected pixels runs across different settings. aborted runs retain observed progress; incomplete API runs have no score. a development snapshot rather than a ranking. no pixels run has reached the flag yet.
recorded runs play in pane a; other model names open their action trace