Multi-scale World Models Enhance Embodied Agents in Dynamic Environments.

Embodied AI agents operating in complex, real-world scenarios require sophisticated methods to reason across different levels of abstraction and adapt their knowledge as conditions evolve. Traditional Mixture of Experts (MoE) approaches often struggle with this, lacking a clear mechanism to select relevant scales for updates or to manage the varying rates at which knowledge at different scales becomes obsolete. A new framework, MuSix, tackles these issues by introducing a scale-aware world model mixture and an evolutionary adaptation process. It employs a two-stage routing system that uses "experiential distance" to determine the appropriate scale for knowledge retrieval, allowing for targeted updates. Furthermore, MuSix implements scale-dependent forgetting rates, ensuring that low-level, rapidly changing information is refreshed frequently, while high-level, more stable abstractions persist. This hierarchical approach, combined with gated inter-scale transfer, helps maintain coherence across the agent's knowledge base, leading to improved performance in multi-scale reasoning and dynamic adaptation tasks.