Neurosymbolic Learning of Interpretable World Models
We introduce Concept-Grounded World Models (CGWMs), a neurosymbolic world model design that instead learns interpretable concept representations.
ProofPaper ↗
Key points
- World models aim to learn compact representations of an environment from observations, enabling agents to predict future states and plan their behaviour.
- Furthermore, we directly use this knowledge by formulating planning as differentiable reasoning, optimising the agent's actions by gradient ascent to satisfy a goal.
- We experimentally compare CGWM with a purely neural baseline and a neurosymbolic baseline on two game environments.
- We find that CGWMs ground concepts much more accurately, even over long horizons, resulting in significantly more robust planning both in and out of distribution.
Sources (1)
- [1]Neurosymbolic Learning of Interpretable World ModelsarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 14, 10:38 AM
We introduce Concept-Grounded World Models (CGWMs), a neurosymbolic world model design that instead learns interpretable concept representations.
World models aim to learn compact representations of an environment from observations, enabling agents to predict future states and plan their behaviour.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 14, 2026Towards Agentic Robots: Model, Data, Environment, and Harness
- Oct 14, 2026SafeDriveVLA: Navigation-Conditioned World Model Dreaming for Conflict-Aware End-to-End Autonomous Driving
- Oct 13, 2026Dyna-2: General Purpose Physical Agents with Hour-Long Autonomy — Jason Ma, Dyna Robotics
- Oct 13, 2026Gemini 4 Argon for Robotics
- Oct 13, 2026WorldMemBench Evaluates Static and Dynamic Memory in Video World Models
- Oct 13, 2026WAVE-0: Bridging VLA Policies and WAMs with Lightweight Latent Future Prediction