DreamWorld: Unified World Modeling in Video Generation

arXiv 26.3 2026 60.5 method

TLDR

DreamWorld integrates multiple world knowledge dimensions into video generation via joint modeling, improving world consistency.

Reasoning

Strengths: novel unified framework addressing limitations of single knowledge alignment, proposes CCA and inner-guidance to stabilize training. Weaknesses: limited evaluation details in abstract (only VBench score), no mention of ablation or comparison to other world model methods.

Read-first score

Read-first score 60.5, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 39.

Recency 8%

100

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

Methodology quality 25%

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

Topical relevance 42%

55.7

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%

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

Field roles

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 320.

Keyword Scores

world model

video world model

generative world model

world dynamics prediction

world simulator

interactive world model

model-based reinforcement learning world model

Deep Analysis

Innovations

Joint World Modeling Paradigm that integrates complementary world knowledge (physical commonsense, 3D, temporal consistency) into video generators by jointly predicting video pixels and features from foundation models
Consistent Constraint Annealing (CCA) to progressively regulate world-level constraints during training
Multi-Source Inner-Guidance to enforce learned world priors at inference

Methodology

DreamWorld is a unified framework that integrates multiple heterogeneous world knowledge dimensions (physical commonsense, 3D, temporal) via joint prediction of video pixels and features from foundation models. It employs Consistent Constraint Annealing (CCA) during training to mitigate visual instability and temporal flickering, and Multi-Source Inner-Guidance at inference to enforce learned world priors. The model is evaluated against the Wan2.1 baseline using the VBench metric.

Key Results

DreamWorld improves world consistency, outperforming Wan2.1 by 2.26 points on VBench.

Resources

Paper