‘Code as Worlds’ Enables AI to Discover Executable World Models for Physical Reasoning
agents reasoning
| Source: HF Papers | Original article
A team of researchers has unveiled “Code-as-World,” a new paradigm that treats physical environments as executable code rather than static pixel arrays. The approach, described in the paper *Code as Worlds: Agentic Discovery of Executable World Representations for Physical Reasoning*, uses an agentic discovery loop to infer compact, generalizable world models that can be run, evolved, and rendered. By converting visual input into a programmatic description of objects, states, and dynamics, the system can simulate outcomes, generate video, and support embodied tasks that require physical intelligence.
The development addresses a long‑standing gap in vision‑language models, which excel at recognizing and describing scenes but often lack explicit representations of the underlying mechanisms that drive physical change. Code-as-World supplies those mechanisms in the form of executable code, enabling quantitative physical reasoning and more reliable prediction of how objects will interact. Early demonstrations show the model composing and evolving code snippets that capture appearance, motion, and causal relationships, hinting at a route toward more robust robotics and simulation‑driven AI.
What follows will be a test of how well the paradigm scales beyond laboratory benchmarks. Researchers are expected to release broader evaluations on tasks such as manipulation planning, video synthesis, and embodied navigation. Integration with existing agentic AI frameworks—systems that set goals and act autonomously—could amplify the impact, turning perception into a programmable substrate for action. Industry observers will watch for collaborations that bring the code‑based world models into commercial robotics pipelines or virtual‑world creation tools, and for any open‑source releases that allow the community to build on the agentic discovery loop. The coming months should reveal whether executable world representations become a cornerstone of next‑generation physical reasoning.
Sources
Back to AIPULSEN