Reproduce “BlockVLA: Accelerating Autoregressive VLA via Block Diffusion Finetuning” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Ruiheng Wang, Shuanghao Bai, Haoran Zhang, Badong Chen, Xiangyu Xu

BlockVLA is a semi-autoregressive discrete diffusion framework for Vision-Language-Action (VLA) models that accelerates inference by 3.3x over discrete diffusion baselines and 8.0x over autoregressive models while achieving higher success rates, such as 91.7% on the LIBERO benchmark, through block-wise parallel action generation and efficient state reuse.

Key insights

  • A hybrid approach combining block-level autoregression with intra-block parallel diffusion effectively balances global causal coherence, local computational efficiency, and error robustness for VLA models.
  • Training with 'Diffusion Forcing,' which conditions on potentially noisy previous states, leads to a more robust coupling between causal blocks compared to 'Teacher Forcing'.
  • Optimizing block size to encompass multiple timesteps (e.g., two consecutive timesteps) and removing the token shift can further improve performance by capturing inter-step dynamic correlations and streamlining prediction targets.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.