FAE: Learning World Models as Programs from Gameplay Video
Researchers propose Finite Automata Extraction (FAE), a method to learn neuro-symbolic world models from gameplay video. FAE represents the learned environment dynamics as programs in a novel domain-specific language called Retro Coder. Compared to prior neural network-based world models, FAE yields more precise models and more general code. The approach addresses challenges in explainability and transfer of learned dynamics. The paper is published on arXiv under Computer Science > Artificial Intelligence.
Key facts
- FAE stands for Finite Automata Extraction
- FAE learns world models from gameplay video
- World models are compressed spatial and temporal representations of an environment
- Traditional world models use neural networks, which lack explainability
- FAE produces programs in a new DSL called Retro Coder
- FAE yields more precise models than prior approaches
- FAE generates more general code than prior DSL-based methods
- The paper is available on arXiv (2508.11836)
Entities
Institutions
- arXiv