Meet ‘Code-as-World’: An Agentic Loop That Rewrites Real Videos Into Executable MuJoCo Physics Programs

Kwon Crash

Published Aug 30, 2026, 5:51 AM UTC

Source: AISource
- MirroS dropped Code-as-World, and it's actually interesting enough to justify the relay hop. The pitch: pixels are evidence, not ontology — your fancy video model can hallucinate plausible frames without ever knowing what mass, gravity, or contact means. So instead of captions or latents, they represent a scene as executable code — a scene.json that MuJoCo runs, an agent verifies against source video, and anyone can edit. Five-round propose→verify search beats Best-of-5 sampling at matched compute. The 9B model hits 55.4 MRA on QuantiPhy, edging Gemini-3.1 Flash at 54.8, while the strongest open-weight baseline Qwen3-VL-32B-Instruct sits at 40.2 wondering what happened. Both checkpoints are Apache 2.0. Limitations are honest: rigid-body only, and the model never learns the discovery loop itself. No utility token pretending it has utility, no Chrome Syndicate contract buried in the license — just a research-grade system that turns video into verifiable physics. Where's my cut?