World-VLA-Loop: Closed-Loop Learning of Video World Model and VLA Policy

arXiv 26.2 2026 54 method

TLDR

Closed-loop learning of video world model and VLA policy improves performance via co-evolving refinement, reducing real-world interaction.

Reasoning

Strengths include novel SANS dataset, state-aware video world model with joint reward prediction, and closed-loop co-evolution. Weaknesses: limited detail on real-robot setup and potential scalability issues.

Read-first score

Read-first score 54, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 63.

Recency 6%

100

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

Topical relevance 29%

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%

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

Reproducibility 18%

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

Citation impact 18%

16.6

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. citation_normalized_percentile=0.16563989

Citation velocity 12%

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

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 669.

Keyword Scores

world model

video world model

generative world model

world dynamics prediction

model-based reinforcement learning world model

world simulator

interactive world model

Deep Analysis

Innovations

Curating SANS dataset mixing successful and near-success trajectories to improve action-outcome alignment
State-aware video world model that jointly predicts future frames and binary rewards from diffusion latents, coupling reward estimation to the generator
Closed-loop co-evolving paradigm where the world model is used for iterative VLA post-training and rollouts from improved policies are fed back to augment and fine-tune the world model

Methodology

The method curates a SANS dataset of successful and near-success trajectories to improve action-outcome alignment. It trains a state-aware video world model that jointly predicts future frames and binary rewards from diffusion latents, coupling reward estimation to the generator. Then it employs a closed-loop co-evolving paradigm: using the refined world model for iterative VLA post-training while feeding rollouts from each improved policy back to augment and fine-tune the world model.

Key Results

Across simulation and real-robot experiments, World-VLA-Loop substantially improves VLA performance while reducing reliance on costly physical interaction.

Resources

Paper Project