Mixture of Contexts for Long Video Generation

arXiv 2025 19.7 method

TLDR

Proposes Mixture of Contexts (MoC) for sparse attention in long video generation, achieving near-linear scaling and long-term consistency.

Reasoning

The paper presents a novel attention routing mechanism for long-context video generation, but lacks explicit real-world evaluation or comparison to baselines. The abstract does not discuss world models or dynamics prediction.

Read-first score

Read-first score 19.7, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 0.

Recency 8%

86.7

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

Reproducibility 25%

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

Methodology quality 25%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=none

Topical relevance 42%

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

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 31.

Keyword Scores

world model

world simulator

generative world model

interactive world model

video world model

world dynamics prediction

model-based reinforcement learning world model

Resources

Paper