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

Code-as-World-VL-9B scores 55.4 MRA on QuantiPhy-validation, above Gemini-3.1 Flash at 54.8. MirroS's Code-as-World paradigm represents scenes as executable code — a scene.json that MuJoCo runs and verifies against source video — rather than pixels or captions. An agentic loop recovers these programs from footage in up to five rounds, and verified worlds provide training data with exact physical labels. MirroS released Code-as-World-VL-4B and Code-as-World-VL-9B, fine-tuned from Qwen3.5-4B and Qwen3.5-9B, under Apache 2.0.