MazeBench leaderboard

MazeBench (MazeBench: visual spatial reasoning in a 3D open world) is an agentic benchmark published by MazeBench Results. Tests whether a coding agent can explore a 256-room 3D puzzle world and collect its 100 hidden gems, using box-pushing puzzles, camera rotation, and long-horizon planning. Its evaluation set is described by the source as 100 hidden gems across 256 rooms in one open world. The source rates its difficulty as long-horizon spatial planning, with single puzzles that can take over one hundred moves.

Each model's result is published by MazeBench Results as gems collected out of 100, with rooms visited as a second score. ModelScale prints that value to 0 decimal places, exactly as published, and never converts or rescales it. This page ranks the 43 models that carry a published value, highest value first; a model the source never scored is left out rather than ranked at zero. The source does not state whether a higher value is the better result, so neither does this page.