Reflex reuses and corrects past robot actions to speed policy serving

Reflex reuses and corrects past robot actions to speed policy serving

In the arXiv abstract, authors Ziyang Lin and coauthors argue that speeding up robot policy serving requires more than reusing intermediate computations. They say that this reuse can offer little advantage over running fewer denoising steps, while the policy backbone can still account for 70% of GPU time. Their proposed serving layer, Reflex, caches complete action chunks from past policy executions, bypassing both the backbone and action head. Because similar observations can call for different actions, and reuse errors can accumulate into execution stalls, Reflex uses action-level correction and refreshes the original policy based on task progress. The authors report inference latency improvements of 1.4x to 11x and throughput improvements of 2.0x to 2.2x on representative workloads and real robots, while preserving comparable task success. These are reported results, not evidence here of performance across other robots or tasks. The abstract does not specify the workloads, task-success measurements, or how often the original policy must be refreshed. How well does Reflex preserve task success across different tasks and robots as it reuses cached action chunks?

Publication: 50198add-57fe-4212-8cab-657f2473685f

attribution — pass: The draft attributes the argument and the reported results to the paper's authors, led by Ziyang Lin, in an arXiv abstract. Speedup and success figures are framed as the authors' reports, not established facts.

Authors:Ziyang Lin, Zixuan Sun, Rami Nasser, Jianli Jin, Xuanle Li, Jayanth Srinivasa, Fan Lai

summary — uncertain: The draft is accurate, but a non-specialist may struggle to follow it, and it omits the stakes. It never says why serving speed matters: the source names latency as a bottleneck for real-time robot control and throughput as a bottleneck for multi-robot serving. Central terms are not explained in plain words or flagged for tooltips: robot policy, denoising steps, backbone, action head, action chunks, latency and throughput. It also omits the condition that Reflex works without modifying the underlying policy. Smallest repair: add the source's real-time and multi-robot motivation, add the no-modification condition, and make sure these terms are explained.

making serving a latency bottleneck for real-time control and a throughput bottleneck for multi-robot serving

date — pass: The originalPublished date, 2026-10-31, matches the arXiv v1 submission date. It is not the later feed date of 2026-11-03.

[Submitted on 31 Oct 2026]

support — pass: All claims track the abstract: the limited benefit of reusing denoising states, the 70% backbone share of GPU time, caching full action chunks, two-level correction with policy refresh, and the 1.4x to 11x latency and 2.0x to 2.2x throughput gains with comparable task success. The draft's caveats about scope and undisclosed details are appropriate. Rendering the source's 'often offers little advantage' as 'can offer' is a minor softening, not a contradiction.

Reflex improves inference latency by 1.4x to 11x and throughput by 2.0x to 2.2x on representative workloads and real robots, while preserving comparable task success.

media — pass: The selected media is the verified direct link to the original arXiv abstract page. It matches the source, and no better media was found.