PapersSelected

StarWM Routes Attention for Robust World Model Training

StarWM: Self-Supervised Trained Attention Routing for Robust World Models

Why we picked it

The technique targets a real failure mode: reconstruction-based world models waste capacity on pixel area instead of dynamics relevance, so agents get distracted by background video. Builders training RL agents or world models can borrow the attention-routed reconstruction pattern to keep representations focused on task-relevant content. It is a paper-level method benchmarked on DeepMind Control, comparable against both reconstruction-based and reconstruction-free baselines.

Researchers propose StarWM, a world model that uses a cross-attention module trained on self-supervised dynamics to decide where reconstruction applies.

Body

A dual-stream decoder restricts reconstruction to attended regions, with stop-gradient barriers separating the two objectives.

On DeepMind Control with dynamic video backgrounds, the default reward-free StarWM achieves the strongest performance under random-frame distractors and outperforms reconstruction-based baselines under sequential video.

Its reward-augmented variant matches or exceeds reconstruction-free methods, achieving the highest overall return across all distractor regimes.

Mechanistic probing shows it preserves state attributes with near-perfect fidelity through long-horizon imagination while discarding distractors.

Original · arXiv: Google DeepMind

StarWM: Self-Supervised Trained Attention Routing for Robust World Models

A robust world model must strike the balance between faithfully capturing environmental dynamics and abstracting away from irrelevant content. While reconstruction-based world models ensure faithful supervision, they misallocate representational capacity by pixel area rather than dynamics relevance for visual tasks, which can cause task-irrelevant content to dominate the learned representation. Alternatively, reconstruction-free methods avoid this bias but risk discarding possibly relevant information. We propose StarWM, which uses a cross-attention module trained on self-supervised dynamics to decide where reconstruction applies. A dual-stream decoder then restricts reconstruction to the attended regions, with stop-gradient barriers preventing interference between the two objectives. These components allows reconstruction to supervise the visual content of attended regions without contaminating the latent with non-predictive information. On DeepMind Control with dynamic video backgrounds, default (reward-free) StarWM achieves the strongest performance under random-frame distractors and substantially outperforms reconstruction-based baselines under sequential video. In addition, its reward-augmented variant matches or exceeds reconstruction-free methods on sequential video, achieving the highest overall return across all distractor regimes. Mechanistic probing confirms StarWM preserves state attributes with near-perfect fidelity through long-horizon imagination while systematically discarding distractors.

Industry Coordinates & Source Verification

This report is categorized under AITOP's "Papers" desk. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows. Key entities and technical topics under continuous tracking include StarWM, world models, DeepMind Control, self-supervised learning.

Primary source verification: published directly by arXiv: Google DeepMind with first-party authority.