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

Learning World Models for Interactive Video Generation

arXiv 25.5 2025 42.4 method

TLDR

Proposes VRAG to reduce compounding errors and improve spatiotemporal consistency in interactive video world models.

Reasoning

Strengths: Identifies key challenges (compounding errors, memory) and proposes a novel retrieval-augmented approach (VRAG) with global state conditioning. Weaknesses: Limited experimental details in abstract; claims about irreducible errors require further validation.

Read-first score

Read-first score 42.4, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 51.

Recency 6%
86.7

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

Topical relevance 29%
72.9

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%
60

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

Reproducibility 18%
30

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=none

Citation impact 18%
0

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. cited_by_count=0

Citation velocity 12%
0

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

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 402.

Keyword Scores

world model
10
interactive world model
10
video world model
10
generative world model
8
world dynamics prediction
6
world simulator
5
model-based reinforcement learning world model
2

Deep Analysis

Innovations

  • Enhancement of image-to-video models with interactive capabilities via action conditioning and autoregressive framework
  • Revelation that compounding error is inherently irreducible in autoregressive video generation
  • Proposal of video retrieval augmented generation (VRAG) with explicit global state conditioning
  • Demonstration that naive autoregressive generation with extended context windows and retrieval-augmented generation are less effective due to limited in-context learning capabilities of current video models
  • Establishment of a comprehensive benchmark for improving video generation models with internal world modeling capabilities

Methodology

The paper enhances image-to-video models with action conditioning and an autoregressive framework to enable interactivity. It proposes video retrieval augmented generation (VRAG) with explicit global state conditioning to address compounding errors and memory limitations. The approach is evaluated against naive autoregressive generation with extended context windows and retrieval-augmented generation, using long video generation tasks to measure compounding errors and spatiotemporal consistency.

Key Results

VRAG significantly reduces long-term compounding errors and increases spatiotemporal consistency of world models, while naive autoregressive generation with extended context windows and retrieval-augmented generation prove less effective due to limited in-context learning capabilities of current video models.

Limitations

  • Compounding errors are inherently irreducible in autoregressive video generation, meaning VRAG can only reduce but not eliminate them.

Resources

Paper

Tags