Reflex

Reflex is a real-time streaming inference framework for flow-matching Vision-Language-Action (VLA) models, accepted at ICML 2026. It enables continuous real-time streaming inference by exploiting the Timestep-Invariance Property: perception encoders (VLM backbone) are independent of the denoising timestep, while the lightweight action expert and flow head are timestep-dependent. By parallelizing vision encoding and action generation, Reflex achieves a 2.58x inference speedup and enables 50Hz stable streaming with 54% reaction latency reduction on the LIBERO and Kinetix benchmarks. The framework targets architectures with separate vision encoders and action experts (e.g., Pi0.5 with PaliGemma backbone), re-architecting the serving loop into decoupled streams with specialized caching, normalization, and pipelining optimizations. Reflex includes future-state prediction for async alignment and is deployable on real robot systems. It maintains numerical stability for long-horizon streaming and enables high-frequency control without performance loss.