Code-as-World agents top QuantiPhy, beat proprietary VLMs
TL;DR
- A vision-language model called Code-as-World-VL reaches state-of-the-art on the QuantiPhy physical-reasoning benchmark and surpasses leading proprietary models.
- The paradigm represents physical worlds as executable code covering composition, dynamics, and visual appearance, then verifies hypotheses by simulation and rendering.
- An agentic discovery loop proposes, executes, renders, verifies, and refines candidate world programs, a process the paper frames as inspired by abductive reasoning.
A vision-language model trained on supervision from verified physics simulations reaches state-of-the-art performance on the QuantiPhy benchmark and surpasses leading proprietary models, according to a new arxiv paper introducing a paradigm the authors call Code-as-World.
The core move is to treat a physical scene as executable code rather than as pixels or captions. "By expressing physical composition, dynamic evolution, and visual appearance as executable code," the paper argues, the framework "provides a compact, quantitatively grounded, and controllable abstraction of the physical world." To build such representations from natural-language descriptions or real videos, an agent proposes candidate world programs, executes and renders them, verifies against the observation, and iteratively refines. The authors describe the loop as "inspired by abductive reasoning."
Those verified worlds are then reused as a training signal. The paper's claim is that they can serve as "scalable physical supervision for training vision-language models on quantitative physical reasoning," and that the resulting model, Code-as-World-VL, surpasses proprietary systems on QuantiPhy.
The abstract publishes no per-benchmark score and does not name which proprietary models were in the comparison.
Originally reported by paper
Read the original article →Original headline: Code as Worlds: Agents That Discover Executable Physical World Models Hit SOTA on QuantiPhy, Beat Frontier Models