DriveVLA-W0: World Models Amplify Data Scaling Law in Autonomous Driving

arXiv 25.10 2025 61.7 method, application

TLDR

DriveVLA-W0 uses world modeling to predict future images, addressing supervision deficit in VLA models and amplifying data scaling law in autonomous driving.

Reasoning

Strengths: novel paradigm combining world modeling with VLA to leverage self-supervised signals, demonstrated on multiple benchmarks with scaling law amplification. Weaknesses: limited to specific VLA archetypes, reliance on large in-house dataset may limit reproducibility, no explicit comparison to other world model methods.

Read-first score

Read-first score 61.7, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 42.

Recency 8%

86.7

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

Methodology quality 25%

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

Topical relevance 42%

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=dataset

Field roles

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 179.

Keyword Scores

world model

generative world model

world dynamics prediction

video world model

world simulator

interactive world model

model-based reinforcement learning world model

Deep Analysis

Innovations

World modeling to predict future images as dense self-supervised signal to address supervision deficit in VLA models
Two instantiations: autoregressive world model for discrete visual tokens and diffusion world model for continuous visual features
Lightweight action expert for real-time deployment
Demonstrates amplification of data scaling law, with performance gains accelerating as training dataset size increases

Methodology

DriveVLA-W0 employs world modeling to predict future images, generating dense self-supervised signals that compel the model to learn underlying driving dynamics. It instantiates this paradigm for two VLA archetypes: an autoregressive world model for discrete visual tokens and a diffusion world model for continuous visual features. A lightweight action expert is introduced to address inference latency for real-time deployment. The model is evaluated on NAVSIM v1/v2 benchmark and a 680x larger in-house dataset, comparing against BEV and VLA baselines.

Key Results

DriveVLA-W0 significantly outperforms BEV and VLA baselines on NAVSIM v1/v2 and the in-house dataset. Crucially, it amplifies the data scaling law, showing that performance gains accelerate as the training dataset size increases.

Resources

Paper