UNDER REVIEW

Making Computation Count:
Speculative State Recycling for Flow-Based VLAs

Anonymous Authors

Paper Code

Overview

Speculative State Recycling overview: speculative inference trade-off, semantic memory and proposal recovery, and real-world execution.

A Flow-VLA speculative system that recycles computed states to improve the quality–efficiency trade-off

Output–cost analysis

Measure speculative efficiency by executable action yield versus complete inference cost; discarded intermediate states are wasted compute.

Speculative State Recycling

Recycle retained semantic states for lightweight prediction and context-dependent refresh, and rejected proposals for short fresh-conditioned generation.

Simulation and real-robot evaluation

LIBERO: nearly matches compiled π0.5 at 1.89× per-action speedup. LIBERO-Plus: +5.77 vs FLASH at +0.15 ms/action. Real Panda: better conveyor, whack-a-mole, and timed pick-and-place.

Abstract

Speculative inference accelerates flow-based vision-language-action models (Flow-VLAs) by bypassing vision-language model (VLM) prefill and full flow integration on speculative rounds. However, lightweight drafting and approximate verification can compromise action quality, while additional computation to improve reliability can erode the resulting speedup. Moreover, rejection can erase these savings by triggering full-path inference after the draft and verification costs have already been incurred. These limitations point to a broader inefficiency: speculative states are often discarded once they are no longer directly usable, even though they may still contain useful predictive information. We propose Speculative State Recycling (SSR), which improves this quality–efficiency trade-off by recycling previously computed states across changes in conditioning. Before refresh, retained deep VLM features serve as semantic memory for a shallow predictor, augmenting the current observation with prior high-level representations. The same shallow computation also produces a draft-independent Context Risk estimate of how much refreshing the conditioning would change the full model’s actions, guiding execution and refresh. When refresh is needed, saved proposals initialize short fresh-conditioned Action Expert integration, reducing regeneration from noise. On LIBERO, SSR with Triton nearly matches the success rate of full-generation π0.5 with compiled inference while achieving a 1.89× speedup per executed action and 74.1% fewer full AE calls per executed action. On LIBERO-Plus, SSR improves mean success by 5.77 points over FLASH at only 0.15 ms/action additional cost.

Method

SSR inference framework with memory-conditioned joint prediction and proposal recycling after rejection.
Speculative State Recycling inference framework. Retained VLM features support memory-conditioned joint prediction of action proposals and Context Risk, which together with Action verification determines the executable prefix and refresh decision. After refresh, rejected proposals initialize short fresh-conditioned recovery, with full generation as fallback.

1 Semantic memory — reuse stale VLM features

Retained deep features condition a shallow joint drafter, so later rounds skip another full prefill.

2 Context Risk — decide the executable prefix

The same shallow pass estimates how much a refresh would change the teacher actions, and triggers refresh when needed.

3 Proposal recovery — start from the rejected draft

After refresh, the rejected proposal initializes short fresh-conditioned Action Expert integration instead of noise.

Real-Robot Videos

Timed four-object pick-and-place (30s)

Pick and place the mango, corn, sponge, and brush within 30 seconds.

π0.5 · fail (44s)

SSR · success (27s)

Conveyor-belt mango grasping

Low / mid / high belt speeds. π0.5 succeeds only at low speed; SSR succeeds at all three.

Low

π0.5 · success

SSR · success

Mid

π0.5 · fail

SSR · success

High

π0.5 · fail

SSR · success

Whack-a-mole

Low / high game speeds. Scores are the hit counts shown on the board.

Low

π0.5 · score 4

SSR · score 11

High

π0.5 · score 2

SSR · score 6

BibTeX

@article{anonymous2027ssr,
  title={Making Computation Count: Speculative State Recycling for Flow-Based VLAs},
  author={Anonymous},
  journal={OpenReview preprint},
  year={2027}
}