WorldSolver: Can LLM Agents Simulate the Physical Dynamics via Solver Generation?
arXiv cs.AI · 7 Oct 2026, 03:26 AEST
LLM-based agents are making strides in scientific and engineering problem solving, particularly in physics simulation. The introduction of WorldSolver benchmarks the ability of these agents to generate solvers for complex physical phenomena.
Key facts
- WorldSolver is a benchmark of 168 simulation tasks derived from physical phenomena in 61 classic computer graphics papers, spanning 7 physical domains.
- Each task contains a code scaffold that provides a fixed simulation environment for the scene, with the solver implementation left for the agent to complete.
- Experiments on frontier agents reveal that producing executable solvers is difficult itself, and satisfying visual and physical correctness is even harder.
- GPT-5.6-Sol and Claude-Opus-5 perform comparatively better than the other evaluated agents, yet achieve overall scores of only 48.7% and 46.7%, respectively.
Read aloud
L L M-based agents are making strides in scientific and engineering problem solving, particularly in physics simulation. The introduction of WorldSolver benchmarks the ability of these agents to generate solvers for complex physical phenomena. WorldSolver is a benchmark of one hundred sixty eight simulation tasks derived from physical phenomena in sixty one classic computer graphics papers, spanning seven physical domains. Each task contains a code scaffold that provides a fixed simulation environment for the scene, with the solver implementation left for the agent to complete. Experiments on frontier agents reveal that producing executable solvers is difficult itself, and satisfying visual and physical correctness is even harder. G P T five.six-Sol and Claude-Opus-five perform comparatively better than the other evaluated agents, yet achieve overall scores of only forty eight point seven percent and forty six point seven percent, respectively.
Sources in this cluster
- WorldSolver: Can LLM Agents Simulate the Physical Dynamics via Solver Generation?
arXiv cs.AI · 7 Oct 2026, 03:26 AEST