MuSix: Multi-scale world-model mixture for embodied agents

MuSix, a multi-scale mixture-of-world-models framework for embodied agents, was submitted to arXiv on 1 Jul 2026 under arXiv:2607.00457 and is listed as "Accepted at ECCV 2026" in the paper comments. The paper, by Jinwoo Jang, Daniel J. Rho, Sihyung Yoon, Hyunsuk Cho and Honguk Woo, introduces a two-stage routing scheme and scale-dependent forgetting rates to handle evolving environments.

What is MuSix and how does it work?

MuSix is a framework that composes world models across continuous scales and adapts them at different rates. The core is a two-stage routing mechanism: a meta-router first maps experiential distance, a measure of situational novelty inspired by Construal Level Theory, to a weight over continuous scale space; then per-scale base routers select world models within the chosen scale. For adaptation, MuSix applies scale-dependent forgetting rates so low-scale knowledge refreshes rapidly while high-scale abstractions persist, and it uses gated inter-scale transfer to maintain coherence across the hierarchy.

The paper frames two practical challenges when applying Mixture of Experts to embodied agents: routing that lacks an explicit notion of scale, and a uniform update policy that cannot match different rates at which knowledge at each scale becomes outdated. MuSix addresses both with its scale-aware routing and evolution mechanisms.

How was MuSix evaluated and what were the results?

MuSix was evaluated on embodied-agent benchmarks, specifically EmbodiedBench and HAZARD, where it improved over state-of-the-art baselines on multi-scale reasoning and dynamic adaptation. The authors report improvements relative to modern baselines on those tasks, positioning MuSix as more effective in environments that change at different temporal or spatial scales.

The paper is 15 pages long and includes experimental details and comparisons. It highlights experiential distance as the grounding signal for scale selection and presents gated inter-scale transfer as the mechanism that preserves coherence while allowing targeted updates across scales.

Why it matters

Embedding scale-awareness into world-model mixtures targets a real gap for agents operating in the real world: different layers of knowledge become stale at different speeds. By combining a continuous scale selection with per-scale update rates, MuSix lets agents refresh short-term, situational models quickly while keeping long-term abstractions stable. That separation of update dynamics matters for embodied systems that must both react to immediate novelty and retain broad, slow-changing structure.

MuSix also ties its routing signal to an interpretable quantity, experiential distance, which roots routing decisions in measurable novelty rather than opaque gating alone. This makes targeted adaptation and controlled inter-scale transfer operational rather than heuristic.

What to watch

Watch for the ECCV 2026 proceedings where the paper was accepted and for the paper's DOI: the arXiv page notes an arXiv-issued DOI via DataCite is pending registration. Also look for code, data or supplementary material linked from the authors' arXiv entry or conference page to reproduce the EmbodiedBench and HAZARD results.

Authors: Jinwoo Jang, Daniel J. Rho, Sihyung Yoon, Hyunsuk Cho, Honguk Woo. arXiv identifier: arXiv:2607.00457. Submission date: 1 Jul 2026. Comment: Accepted at ECCV 2026. 15 pages.