Venue: ICLR 2026 · Authors: Jinwoo Jang, Minjong Yoo, Sihyung Yoon, Honguk Woo (Sungkyunkwan University) · arXiv: 2601.22647 · Category: World model for actions · Trend tag: test-time-adaptive world models.

flowchart LR
  Obs[Unseen-domain<br/>observation features] --> Route[Multi-granular<br/>prototype routing<br/>object to scene level]
  Route --> Refine[Test-time refinement<br/>align features to prototypes]
  Refine --> Mix[Mixture over<br/>world models / experts]
  Mix --> Aug[Distilled mixture<br/>augmentation<br/>build new models few-shot]
  Aug --> Pred[World-model prediction<br/>for reasoning / planning]
  Pred --> Act[Embodied action]

Language-model-based embodied agents need accurate, flexible world models to reason and act in dynamic environments. Conventional Mixture-of-Experts (MoE) routing is fixed once deployed, so it generalizes poorly to unseen domains.

TMoW (Test-time Mixture of World Models) extends MoE so the routing function over world models is updated at test time, letting agents recombine existing models and integrate new ones for continual adaptation. Three mechanisms:

• Multi-granular prototype-based routing — adapts mixtures across object-level to scene-level similarities.
• Test-time refinement — aligns unseen-domain features with prototypes during inference.
• Distilled mixture-based augmentation — efficiently constructs new models from few-shot data and existing prototypes.

• Evaluated on VirtualHome, ALFWorld, and RLBench.
• Reports strong performance in both zero-shot adaptation and few-shot expansion scenarios. (Paper does not surface headline numeric scores in the abstract; specific figures omitted here pending the results tables.)

Reframes world-model selection as a test-time-adaptive routing problem rather than a frozen mixture, enabling embodied agents to continually adapt and absorb new world models in changing environments — a notable step in making world-model-driven reasoning robust to domain shift.

• arXiv: https://arxiv.org/abs/2601.22647
• OpenReview: https://openreview.net/forum?id=LQD1MrnbxH

• ICLR 2026 Survey
• VLA Architectures review (world-model category)

← Back to ICLR-2026