A model-based reinforcement learning framework that effectively extracts task-relevant information from noisy observations by leveraging conditional dependence.