FlashVLA: Streaming Action Decoding for Fast and Asynchronous VLA Inference

Authors/institution: Zekai Li and Zhijian Liu, UC San Diego; Jiaming Tang, MIT. Status: arXiv preprint, 27 August 2026. Problem: high latency and temporal mismatch in robot VLA control. Contribution: a staggered action buffer and chunk-wise causal attention that emits one executable chunk per inference step. Practical significance: potentially smoother real-time control. Limitations: requires fine-tuning; results are limited to the authors’ simulated and real-world setups. Paper.