DishwasherBench – Can Frontier Models Stack a Dishwasher?

DishwasherBench is a physical simulation game designed to benchmark language models. The game, Spotless, can be played in a web browser using 3D models of real dishwashers, such as the Bosch SMV4HAX40G. Players load the rigid-body physics model into the dishwasher and simulate a wash, scoring points for every clean item. Currently, the simulation only scores whether cups, glasses, and pots come out mouth down, but it does not yet assess whether plates and cutlery are stacked for proper cleaning.

Version 0.1.0 is an early release, and the simulation is calibrated on one test wash of the real machine. The result shows whether a player packed well, but it does not show if plates and cutlery were stacked for clean washing. The simulation runs on a MacBook Pro (Apple M3 Pro, 18 GB) under macOS 27, and the results are repeatable on that machine. However, the results may vary on different operating systems, webGPUs, and browser configurations.

To play the game, simply load the first level and place pieces in the dishwasher. The game provides guidance on where to place pieces, but some features, such as placing pieces by pointing, are still in development. The simulation checks for soil and drying, but it does not model how spots and film on glass settle or how items move by the jets. The game is scored using Eco 50 detergent on the Finish Powerball Ultimate, and the levels are rated based on what is on the counter, not on actual scores.