Paper notes
Code2World: A GUI World Model via Renderable Code Generation
First arXiv submission: · arXiv
Predicts a GUI's next visual state by generating renderable code, and evaluates whether the predicted outcomes help mobile-agent navigation.
Contribution and setting
Constructs paired screen/action/code data and trains with rendered visual feedback to balance appearance and action consistency.
- Model output
- Renderable UI code, converted into the predicted next screen.
- Action interface
- A current GUI observation and a proposed interaction, represented for the model with action descriptions and visual cues.
- State and memory
- Prediction conditions on the current screen and action; renderable structure represents the output interface.
Relationship to NeuralOS
Code2World cites NeuralOS in Section 2.2 and its bibliography.
Scope and limitations
Uses generated code and a browser to predict screens for Android navigation.
Related terms: Code2World, GUI world model, renderable code, next-screen prediction, AndroidCode.
Connections between methods
Direct frame generation and renderable-code prediction
NeuralOS predicts desktop frames using learned state and a diffusion renderer. Code2World predicts a UI through code that is then rendered into a screen, and tests its usefulness for mobile navigation. Code2World cites NeuralOS in Section 2.2.
Cite this paper
ArXiv preprint
arXiv:2602.09856, initial submission 2026-02-10.