WorldSolver: Can LLM Agents Simulate the Physical Dynamics via Solver Generation?
WorldSolver tests whether agents can write the missing physics solver, and the best reported scores stay below 50%.
The benchmark has 168 simulation tasks drawn from 61 classic computer graphics papers across seven physical domains. Each task gives the agent a fixed scene scaffold and leaves the solver code to be completed. The paper evaluates execution, visual fidelity, and physical plausibility, finding that even runnable solvers are hard to produce and correct dynamics are harder. GPT-5.6-Sol and Claude-Opus-5 lead the tested agents at 48.7% and 46.7% overall. ArXiv · AI/CL/LG's note
The benchmark has 168 simulation tasks drawn from 61 classic computer graphics papers across seven physical domains. Each task gives the agent a fixed scene scaffold and leaves the solver code to be completed. The paper evaluates execution, visual fidelity, and physical plausibility, finding that even runnable solvers are hard to produce and correct dynamics are harder. GPT-5.6-Sol and Claude-Opus-5 lead the tested agents at 48.7% and 46.7% overall. ArXiv · AI/CL/LG's note
score 5