Neurosymbolic World Models for Zero-Shot Task Transfer in Reinforcement Learning

arXiv preprint 2608.17959 presents a neurosymbolic world model formulation aimed at enabling zero-shot task transfer in reinforcement learning. Current model-based reinforcement learning methods learn neural world models that support policy improvement via planning in a latent space, without assuming knowledge of the environment's structure. These models, however, are typically task-dependent, learning uninterpretable latent representations tied to the training task, which impedes generalization to new tasks. The proposed approach decouples observation reconstruction from reward prediction, so that reward prediction depends only on a subset of structured, symbolic components of the latent state. With the reward function defined over a shared symbolic state space, a learned model can adapt to new reward functions without additional environment interactions—i.e., zero-shot. The authors discuss the advantages and challenges of learning such neurosymbolic world models and demonstrate strong generalization properties. The preprint was announced as a new submission on arXiv and focuses on the conceptual framework rather than specific experiments or datasets.