Self-supervised Hierarchical Visual Reasoning with World Model

arXiv 2026 66.4 method

TLDR

ResDreamer proposes a hierarchical world model with residual learning for efficient visual reasoning in RL, achieving state-of-the-art efficiency.

Reasoning

The paper introduces a novel hierarchical residual architecture for world models that is self-supervised and scalable. However, the abstract lacks specific experimental details and does not mention real-world validation, limiting the assessment of its practical impact.

Read-first score

Read-first score 66.4, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 55.

Recency 6%

100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Reproducibility 18%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=True; dataset=False; markers=code,github

Citation impact 18%

80.7

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. citation_normalized_percentile=0.80666209

Topical relevance 29%

78.6

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Methodology quality 18%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=experiment

Citation velocity 12%

Citation velocity estimates citations per publication-year to reduce old-paper bias. velocity=0.00

Field roles

FoundationFrontierBridgeReproducibility anchor

Rank sensitivity

Stability: volatile; rank range: 417.

Keyword Scores

world model

world dynamics prediction

model-based reinforcement learning world model

generative world model

interactive world model

video world model

world simulator

Deep Analysis

Innovations

Hierarchical world model where each higher-level layer reconstructs residuals of the layer below, enabling progressive abstraction of world dynamics.
Self-supervised training of reasoning representations inspired by the 'Bitter Lesson', without domain-specific knowledge.
Residual representations from higher layers modulate lower-level predictions, achieving linear cross-layer communication cost scaling.

Methodology

ResDreamer is a hierarchical world model. Each higher-level layer is trained to reconstruct the residuals of the layer below, fostering progressive abstraction of increasingly sophisticated dynamics. The higher-level residual representations modulate lower-level predictions, and the entire model is trained in a purely self-supervised manner, with cross-layer communication costs scaling linearly.

Key Results

ResDreamer achieves state-of-the-art sample efficiency and parameter efficiency in 3D open-world environments with adversarial opponents.

Resources

Paper Code