Multi-scale Mixture of World Models for Embodied Agents in Evolving Environments
06:00 · July 2, 2026 · arXiv cs.AI RSS
Embodied agents operating in the real world require multi-scale reasoning and knowledge adaptation as conditions change. We identify two challenges in applying Mixture of Experts (MoE) to this setting: routing lacks an explicit notion of scale, preventing targeted updates at specific scales, and a uniform update policy cannot accommodate the different rates at which knowledge at each scale becomes outdated. We present MuSix, a framework that addresses both challenges through scale-aware world model mixture and evolution. A two-stage routing mechanism grounds scale selection in experiential distance, a measure of situational novelty inspired by Construal Level Theory: a meta-router first maps this quantity to a weight over continuous scale space, then per-scale base routers select world models within the identified scale. For adaptation, scale-dependent forgetting rates allow low-scale knowledge to refresh rapidly while high-scale abstractions persist, and gated inter-scale transfer maintains coherence across the hierarchy. Experiments on EmbodiedBench and HAZARD show that MuSix improves over state-of-the-art baselines on multi-scale reasoning and dynamic adaptation.
Summary
Embodied agents must reason across multiple temporal and spatial scales while continuously updating their internal models as real-world conditions evolve. Standard Mixture of Experts architectures fall short here because their routers operate without an explicit scale variable and because they apply a single update rule to all experts, ignoring that low-level sensory-motor knowledge becomes obsolete faster than high-level abstractions.
MuSix addresses both limitations by organizing a collection of world models into a scale-aware hierarchy. A two-stage router first computes experiential distance, a scalar measure of situational novelty drawn from Construal Level Theory, and uses a meta-router to produce a distribution over a continuous scale axis. Within each selected scale, dedicated base routers then activate the most relevant world models.
Adaptation follows the same hierarchy: each scale is assigned its own forgetting rate so that fine-grained predictions can be refreshed quickly while coarser representations remain stable. Gated transfer links between scales preserve consistency when lower-level updates occur. On the EmbodiedBench and HAZARD benchmarks the resulting system outperforms prior state-of-the-art agents on tasks that require both multi-scale planning and rapid response to environmental change.
Why it matters
This research is highly relevant for Dutch AI researchers and robotics practitioners developing embodied agents for dynamic environments, such as those in manufacturing, agriculture, or healthcare. The novel MuSix framework offers advanced methodologies for multi-scale reasoning that can directly inform R&D at Dutch technical universities and high-tech enterprises.
