Beyond Single-Speed Reasoning: Coordinating Fast and Slow Dynamics for Efficient World Modeling

Model-based reinforcement learning (MBRL) enables efficient decision-making by learning predictive world models of environment dynamics. Despite recent advances, existing models often struggle to reconcile accurate short-term transitions with coherent long-term planning, especially in partially observable or long-horizon settings. We argue that this limitation often stems from modeling all transitions at a single temporal resolution, which makes it challenging to simultaneously capture fine-grained local dynamics and abstract global structures. To this end, we propose SF-RSSM (Slow-Fast Recurrent State-Space Model), a novel method that decouples short-term and long-term dynamics via a dual-branch design. The fast branch captures short-horizon transitions using residual prediction, while the slow branch models longrange dependencies with a GRU-based recurrent pathway. A distillation mechanism is developed to enable cooperation across timescales, with the slow model providing soft targets to guide the fast model. Additionally, a curiosity module encourages exploration by promoting learning in regions where the fast and slow branches exhibit divergent dynamics. Experiments on CARLA, DMControl and Atari benchmarks show that SF-RSSM outperforms strong baselines in policy performance.