Tag
Code-as-World introduces a paradigm that represents physical environments as executable code to enable quantitative reasoning for vision-language models, using an agentic discovery loop for scalable supervision and achieving state-of-the-art performance.