Training-Free Inference-Budget Reduction in Flow-Matching Vision–Language–Action Models for Robot Manipulation: An Evaluation Against a Prespecified Task-Success Margin

Flow-matching vision–language–action (VLA) models generate robot actions through iterative denoising, making inference latency a deployment constraint. The acceleration methods reviewed here rarely assess task-success retention against a margin fixed before evaluation. We study three weight-preserving runtime levers on SmolVLA and π0.5: denoising-step count, masked camera prefix elimination, and execution horizon. In simulated closed-loop LIBERO evaluations, three prespecified settings each met a five-percentage-point aggregate task-success margin on 800 paired episodes in the original primary analysis. Because the first confirmation block partly overlapped selection, secondary checks used states the selection grid never reached; they met the margin when pooled, though not in every single block. At settings closer to each checkpoint’s reference evaluation, all three also met it on two blocks. Prediction latency fell by 56–58%; the measured amortized cost per control step for the joint SmolVLA setting was about 94% lower under the stated timing boundaries, with similar reductions on an embedded Jetson AGX Thor. A curvature–tolerance screen agreed with one tested Meta-World rejection but failed to distinguish LIBERO suites, supporting only preliminary screening. Runtime budget changes yielded larger latency reductions than the tested pruning configurations. The contribution is a scoped empirical evaluation and a reusable margin-based procedure, rather than a universal budget or screening guarantee.