WorldCache: Content-Aware Caching for Accelerated Video World Models

arXiv 26.3 2026 70.2 method

TLDR

WorldCache accelerates video world models via content-aware caching, achieving 2.3x speedup with 99.4% quality retention.

Reasoning

Strengths include a novel motion-adaptive caching framework that significantly speeds inference while preserving quality, with clear empirical results. Weaknesses are limited evaluation to a single model and benchmark, and lack of discussion on generalization or interactive settings.

Read-first score

Read-first score 70.2, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 43.

Recency 8%

100

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

Reproducibility 25%

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

Topical relevance 42%

61.4

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 25%

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

Field roles

FrontierReproducibility anchor

Rank sensitivity

Stability: volatile; rank range: 191.

Keyword Scores

video world model

world model

world dynamics prediction

generative world model

world simulator

interactive world model

model-based reinforcement learning world model

Deep Analysis

Innovations

Motion-adaptive thresholds for feature reuse
Saliency-weighted drift estimation
Optimal approximation via blending and warping
Phase-aware threshold scheduling across diffusion steps

Methodology

WorldCache is a Perception-Constrained Dynamical Caching framework for Diffusion Transformers (DiTs) that improves both when and how to reuse features across denoising steps. It introduces motion-adaptive thresholds, saliency-weighted drift estimation, optimal approximation via blending and warping, and phase-aware threshold scheduling, all without retraining. The method is evaluated on the Cosmos-Predict2.5-2B model using the PAI-Bench dataset.

Key Results

WorldCache achieves a 2.3× inference speedup while preserving 99.4% of baseline quality, substantially outperforming prior training-free caching approaches.

Resources

Paper Code