SolarWM: Open Data and Scalable Training for Long-Horizon Video World Models

arXiv 2026 41.5 method, system

TLDR

SolarWM presents an open data engine and backbone-native adaptation framework for training interactive video world models across heterogeneous datasets, enabling long-horizon real-time rollouts.

Reasoning

The paper's strengths include a fully open pipeline, large-scale multi-source data unification, and a unified training recipe across multiple video backbones. However, the abstract lacks quantitative evaluation results or comparisons, so claims about real-time interaction and long-horizon quality are not yet substantiated.

Read-first score

Read-first score 41.5, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 51.

Recency 5%

100

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

Topical relevance 25%

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 15%

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

Reproducibility 15%

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

Graph prestige 15%

5.7

PageRank-style prestige over the local citation graph built from available OpenAlex references. local_pagerank=0.0016

Citation impact 15%

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

Citation velocity 10%

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

Field roles

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 519.

Keyword Scores

video world model

world model

interactive world model

generative world model

world simulator

world dynamics prediction

model-based reinforcement learning world model

Deep Analysis

Innovations

Introduces SolarWM, a fully open foundation for interactive video world models spanning data preparation through long-horizon inference.
Proposes a reconfigurable multi-source data engine that converts 1.43 million canonical clips from 10 datasets into a unified frame-aligned contract and decouples source processing from mixture construction.
Introduces a backbone-native adaptation framework with shared camera-conditioning, training, and inference interfaces across Wan2.2, LTX-2.5, and MiniMax-H3.
Develops a unified three-stage training recipe combining bidirectional adaptation, teacher-forced autoregressive initialization, and distribution matching distillation.
Enables real-time interaction over minute-to-hour rollouts from models trained only on 5-second sequences.
Releases data, pipeline, recipes, weights, and framework to support reproducible and extensible world-model research.

Methodology

SolarWM uses a reconfigurable multi-source data engine to convert 1.43 million clips from 10 datasets into a unified frame-aligned contract covering visual observations, metric camera geometry, captions, quality metadata, selection decisions, and provenance. It instantiates four 5B-33B models based on Wan2.2, LTX-2.5, and MiniMax-H3 under shared camera-conditioning, training, and inference interfaces while preserving native representations and objectives. A three-stage recipe combines bidirectional adaptation, teacher-forced autoregressive initialization, and distribution matching distillation, with training on 5-second sequences.

Key Results

The resulting causal models enable real-time interaction over rollouts ranging from minutes to hours after being trained on only 5-second sequences.

Resources

Paper