The top AI coding-agent setup scored 48.7 out of 100 on a new test of writing physics-simulation code. The WorldSolver preprint, submitted to arXiv on Oct. 6, asks whether AI agents can turn a description of a physical scene into a working solver.
A physics solver is code that advances a simulated system step by step, calculating how its parts move and interact. Researchers derived 168 tasks from 61 computer-graphics papers, across seven areas including fluids, rigid bodies, elastic materials and cloth. Each task gives agents the same fixed scene, physical parameters and renderer, leaving them to write only the solver.
WorldSolver checks whether the code runs, whether its video shows the requested motion, and whether the underlying trajectory passes task-specific physics checks. GPT-5.6-Sol judged 20 sampled frames per video for visual fidelity and also served as the coding agent in the top-scoring setup. The authors compared that judge with expert ratings on 231 of 661 cases that produced videos; the scores correlated at 0.730, meaning higher model ratings tended to coincide with higher expert ratings, not that 73 percent of judgments matched. Separate calculators checked constraints such as momentum balance and contact. Runs that do not execute score zero; completed runs average their visual and physics scores.
That distinction can reveal a plausible-looking mistake. In one Billiard Impulse Chain task, a solver that ran successfully scored 99.5 for visual fidelity but 73.734 for physical plausibility: its simulated momentum change did not match the external forces recorded by the solver.
Across seven tested setups, GPT-5.6-Sol with Codex CLI scored 48.7; Claude-Opus-5 with Claude Code followed at 46.7, and Gemini-3.7-Flash scored 29.3. Each agent had up to 30 minutes to generate a solver and 10 minutes to run it. The agent software setups differed, so the scores compare complete setups rather than models alone. The preprint covers only a subset of physical phenomena and does not establish reliability in deployed simulation work.
Sources
- WorldSolver: Can LLM Agents Simulate the Physical Dynamics via Solver Generation? (arXiv preprint, submitted Oct. 6, 2026)
- Lead image: detail from Figure 3 of WorldSolver by Siru Jiang et al., licensed under CC BY 4.0.