Researchers propose a new type of world model that combines neural networks with symbolic components. This allows for zero-shot task transfer, meaning the model can adapt to new tasks without requiring additional training data. The approach decouples observation reconstruction and reward prediction, enabling more interpretable latent representations. This could improve the generalization properties of model-based reinforcement learning methods. --- Why it matters: This work matters because it addresses a key challenge in AI: adapting to new tasks without extensive retraining. By developing models that can generalize across tasks, researchers can accelerate progress in areas like robotics and autonomous systems. Source: https://arxiv.org/abs/2608.17959