Multi-scale Mixture of World Models for Embodied Agents in Evolving Environments

2026-07-01

AI Summary

This work addresses the challenges faced by existing embodied agents in performing multi-scale reasoning within dynamic environments and the lack of targeted update and adaptive forgetting mechanisms for knowledge at different abstraction levels in mixture-of-experts models. To this end, the paper introduces MuSix, a novel framework that explicitly incorporates scale into the mixture architecture of world models. Drawing on construal level theory to define experiential distance, MuSix employs a two-stage scale-aware routing mechanism. It further integrates scale-dependent forgetting rates and a gated cross-scale knowledge transfer strategy, enabling differentiated updating and co-evolution of multi-granularity knowledge. Experimental results demonstrate that MuSix significantly outperforms current methods on the EmbodiedBench and HAZARD benchmarks, achieving state-of-the-art performance in multi-scale reasoning and dynamic adaptation tasks.

Abstract

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scales, and a uniform update policy cannot accommodate the different rates at which knowledge at each scale becomes outdated.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.

Problem

• multi-scale reasoning
• embodied agents
• knowledge adaptation
• Mixture of Experts
• evolving environments

Innovation

• scale-aware routing
• multi-scale world models
• experiential distance
• scale-dependent forgetting
• gated inter-scale transfer