Neurosymbolic World Models Enable Zero-Shot Task Transfer in AI

State-of-the-art model-based reinforcement learning (RL) often relies on neural world models to learn latent representations of an environment, facilitating policy improvement through planning. However, these neural models typically learn uninterpretable latent states that are highly specific to the training task, making it difficult to generalize them to new tasks without extensive retraining or further interaction with the environment. This research proposes a new formulation for world models called neurosymbolic world models. The core idea is to separate the process of reconstructing observations from the process of predicting rewards. In this new model, reward prediction is made dependent only on a structured, symbolic subset of the overall latent state. This decoupling is crucial because it allows the model to learn more interpretable representations. By focusing reward prediction on symbolic components, these neurosymbolic world models gain the ability to adapt to entirely new reward functions in a "zero-shot" manner. This means they can perform new tasks without any additional environment interactions, provided the new reward function is defined over the same symbolic state space. The study highlights the strong generalization properties of this approach compared to purely neural methods, addressing a significant challenge in making RL systems more flexible and adaptable.