GPUPhysBench: Benchmarking Coding Agents for Correct and Efficient GPU Physics Simulation
Top coding agents could solve every task, but still rarely matched expert GPU speed.
GPUPhysBench tests agents on 50 physics-simulation GPU coding tasks across fluids, deformable solids, and granular materials. In the paper’s single-attempt evaluation, the two strongest model-harness pairs passed all 50 tasks. Speed was the harder problem: even the fastest reached at least 90% of the expert reference runtime on only 22% of tasks, with the biggest gaps in collision detection, constraint solving, and iterative solvers. ArXiv · AI/CL/LG's note
GPUPhysBench tests agents on 50 physics-simulation GPU coding tasks across fluids, deformable solids, and granular materials. In the paper’s single-attempt evaluation, the two strongest model-harness pairs passed all 50 tasks. Speed was the harder problem: even the fastest reached at least 90% of the expert reference runtime on only 22% of tasks, with the biggest gaps in collision detection, constraint solving, and iterative solvers. ArXiv · AI/CL/LG's note
score 6