FlashDrive: A Co-Design Framework for Real-Time VLA Inference in Autonomous Driving

A new arXiv paper (2608.12932) introduces FlashDrive, an algorithm-system co-design framework aimed at reducing the computational cost of Vision-Language-Action (VLA) models for autonomous driving. The authors identify four structural bottlenecks in VLA inference: visual encoding waste on overlapping frames, language-model prefill recomputation, serial generation of low-entropy reasoning tokens, and uniform compute in flow-matching denoising. FlashDrive targets all four stages simultaneously, leveraging temporal overlap for streaming KV-cache reuse, low per-token entropy, and intra-block shortcuts. The paper is categorized as 'new' on arXiv and is available at the provided URL. The framework promises to make end-to-end reasoning feasible for real-time control in autonomous vehicles.