Summary

Reflex introduces streaming inference for vision-language-action models, enabling real-time robot control at much higher control frequencies than standard VLA generation pipelines. By reformulating VLA generation as an incremental streaming process rather than a full-sequence generation step, Reflex decouples inference latency from action-chunk length and allows the robot to begin executing the earliest tokens of an action sequence before generation is complete.

Key Contributions

  • Streaming inference formulation: VLA outputs are consumed incrementally as they are generated, eliminating the full-sequence generation bottleneck
  • Enables real-time VLA control frequencies compatible with high-speed manipulation tasks
  • Accepted at ICML 2026, providing strong peer-reviewed validation of the approach
  • Addresses the fundamental latency-capability tradeoff in deploying large VLA models for reactive robot control

Significance

ICML 2026 acceptance for a core VLA inference efficiency paper signals mainstream recognition of the real-time deployment gap — Reflex offers a principled streaming solution that does not sacrifice model capability for speed, potentially enabling a new class of reactive foundation-model-based robot controllers.