⌁ Awesome World Model Hub Papers · Datasets · Projects
中文 GitHub
← Back to papers

WorldVLA: Towards Autoregressive Action World Model

arXiv 25.6 2025 69.5 method

TLDR

WorldVLA unifies vision-language-action and world models for autoregressive action and image generation, with an attention mask to improve action chunking.

Reasoning

The paper presents a novel integration of action and world models, showing mutual enhancement and addressing autoregressive error propagation with an attention mask. However, the abstract lacks details on real-world evaluation and specific benchmarks, limiting assessment of generalizability.

Read-first score

Read-first score 69.5, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 53.

Recency 8%
86.7

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

Topical relevance 42%
75.7

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%
73

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

Methodology quality 25%
50

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

Field roles

FrontierReproducibility anchor

Rank sensitivity

Stability: volatile; rank range: 94.

Keyword Scores

world model
10
generative world model
9
world dynamics prediction
9
interactive world model
8
video world model
7
world simulator
6
model-based reinforcement learning world model
4

Deep Analysis

Innovations

  • Unifying Vision-Language-Action (VLA) model and world model in a single autoregressive framework for mutual enhancement between action generation and visual prediction.
  • Attention mask strategy that selectively masks prior actions during generation of the current action to mitigate error propagation in autoregressive action chunk generation.

Methodology

WorldVLA integrates a VLA model with a world model, where the world model predicts future images by leveraging both action and image understanding to learn environment physics, and the action model generates subsequent actions based on image observations. The framework is trained and evaluated on tasks involving action chunk generation, using an autoregressive approach with a proposed attention mask to address error propagation.

Key Results

WorldVLA outperforms standalone action and world models, demonstrating mutual enhancement. The attention mask strategy significantly improves performance in the action chunk generation task by reducing error propagation from earlier actions.

Limitations

  • The action model's performance deteriorates when generating sequences of actions autoregressively due to limited generalization capability for action prediction.
  • Error propagation from earlier actions to subsequent ones in autoregressive action generation.

Resources

PaperCode

Tags