LetTokenBurnWorldSolver: Can LLM Agents Simulate the Physical Dynamics via Solver Generation?
A benchmark of 168 tasks evaluating LLM agents' ability to generate physics solvers from descriptions across 7 domains, assessed via execution, visual fidelity, and physical plausibility.
$ npm i -g lettokenburn-cli$ ltb bench info worldsolver
External harness. No Harbor adapter has been generated for this benchmark yet. `ltb bench info` prints the known links.
Runs happen on your machine with your own subscription; only the score summary is pushed. CLI docs
| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| GPT-5.6 Sol | 48.7 | N/A | Independent | medium | Oct 6, 2026 | arxiv.org |
| Claude Opus 5 | 46.7 | N/A | Independent | medium | Oct 6, 2026 | arxiv.org |
| Gemini 3.7 Flash | 29.3 | N/A | Independent | medium | Oct 6, 2026 | arxiv.org |
| DeepSeek V4.1 Flash | 24.6 | N/A | Independent | medium | Oct 6, 2026 | arxiv.org |
| GLM-5.3 | 23.6 | N/A | Independent | medium | Oct 6, 2026 | arxiv.org |
| Qwen3.8 Max | 22.7 | N/A | Independent | medium | Oct 6, 2026 | arxiv.org |
| Kimi K2.7 Code | 17.1 | N/A | Independent | medium | Oct 6, 2026 | arxiv.org |
Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.
WorldSolver comprises 168 simulation tasks derived from 61 classic computer graphics papers, spanning 7 physical domains. Each task provides a phenomenon description and fixed code scaffold, requiring agents to implement the solver logic.
Evaluation combines execution checks, visual fidelity of rendered videos, and physical plausibility of state trajectories. Frontier models like GPT-5.6-Sol achieve only ~48.7%, highlighting the difficulty of agentic solver generation.