⌁ Awesome World Model Hub Papers · Datasets · Projects
中文 GitHub
← Back to papers

Long-Context State-Space Video World Models

arXiv 25.5 2025 70.8 method

TLDR

Proposes a state-space model architecture for video world models that extends temporal memory efficiently, evaluated on Memory Maze and Minecraft.

Reasoning

The paper introduces a novel block-wise SSM scanning scheme to address long-term memory limitations in video diffusion world models, with strong empirical results on spatial retrieval tasks. However, the evaluation is limited to simulated environments, and the abstract does not provide details on real-world deployment or comparisons to model-based RL baselines.

Read-first score

Read-first score 70.8, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 53.

Methodology quality 25%
90

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

Recency 8%
86.7

Uses a gentle age decay so recent papers surface without erasing older foundations. 2025

Topical relevance 42%
75.7

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

Reproducibility 25%
38

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

Field roles

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 53.

Keyword Scores

world model
10
video world model
10
generative world model
8
interactive world model
8
world dynamics prediction
7
world simulator
5
model-based reinforcement learning world model
5

Deep Analysis

Innovations

  • Novel architecture using state-space models (SSMs) for long-term temporal memory in video world models, overcoming computational cost of attention layers
  • Block-wise SSM scanning scheme that strategically trades spatial consistency for extended temporal memory
  • Dense local attention to ensure coherence between consecutive frames

Methodology

The paper proposes a causal video world model that combines state-space models (SSMs) with a block-wise scanning scheme to extend temporal memory efficiently, paired with dense local attention for frame-to-frame coherence. The model is trained on action-conditioned autoregressive frame prediction and evaluated on long-term spatial retrieval and reasoning tasks using Memory Maze and Minecraft datasets.

Key Results

The proposed approach surpasses baselines in preserving long-range memory on Memory Maze and Minecraft datasets while maintaining practical inference speeds suitable for interactive applications.

Limitations

  • Trade-off between spatial consistency and extended temporal memory may degrade fine-grained spatial details
  • Evaluation limited to Memory Maze and Minecraft datasets, leaving generalization to other domains unverified
  • Inference speed is described as 'practical' but not quantitatively compared to baselines or real-time requirements

Resources

Paper

Tags