acceptodds
Under review as a conference paper at ICLR 2027

One Step Is Often Enough: Accelerating VLA Action Heads with Training-Free Marginal-Aligned Warm Starts

Abstract

Vision-language-action models with a flow-matching action head spend several denoising steps on every policy call. We ask how few the head needs at inference, without training, and find that one step is often enough but not always. The same single step keeps or improves success on some tasks and collapses on others. The split is not visible in open loop, where one step returns an estimate of the posterior mean with a small residual of similar size on every task. It is decided in closed loop, where that residual is persistent and integrates in the state until it exceeds the loop's tolerance. A one-step head therefore needs a starting point that carries the executed motion and lies on the marginal it was trained on. Marginal-Aligned Warm Start () provides both without training, by shifting the previously executed chunk, re-noising it to the flow's training marginal, and taking one step. Across models, simulated benchmarks and a real robot, lowers the residual as predicted, recovers most or all of the success that one step from noise loses, and outperforms training-free baselines that spend several evaluations per call. Because the warm start adds little to a one-step call, cutting the steps cuts the time. On the action head runs 9.1 faster at an unchanged suite average.

Primary area:
applications to robotics, autonomy, planning
Keywords:
vision-language-action models, flow matching, efficient inference, training-free acceleration, few-step sampling, warm start, action chunking

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_9x4sOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Marginal-Aligned Warm Start (MAWS), a training-free method for reducing the denoising steps of flow-matching VLA action heads to a single step. The authors study why one-step inference works on some tasks but fails on others, and argue that persistent prediction errors can accumulate during closed-loop execution. MAWS addresses this by shifting the previous action chunk, adding noise according to the flow-matching schedule, and using it to initialize the next prediction.

Critical strengths
  • The paper is easy to follow, with clear and well-designed figures and tables.
  • The proposed method makes sense and is simple and effective.
  • The experiments are quite comprehensive, covering several simulation benchmarks, and real-world robotic tasks.
Critical weaknesses
  • The novelty of the proposed method is relatively limited. Using the previous action chunk as a warm start is not a new idea. Similar approaches have already been explored in previous works, including RTI-DP and Falcon. The main difference here is re-noising the previous chunk according to the flow-matching schedule. While this is a reasonable design, I am not sure it represents a substantial methodological contribution.

  • The experiments do not fully support the proposed motivation. The authors argue that MAWS works by reducing persistent residual errors during closed-loop execution. Although Section 3.2 provide some evidence, I am not convinced that this is the main reason for the performance improvements. Reusing the previous action chunk naturally introduces temporal smoothing, which could also explain the results. I would like to see a more direct comparison with simple temporal ensembling to better understand what the proposed marginal alignment actually contributes.

  • The practical acceleration is less impressive than the action-head numbers suggest. Although MAWS reduces the action-head computation from 10 steps to 1, the end-to-end latency in the real-world experiments only decreases from 67.9 ms to 54.7 ms. This is a relatively modest improvement compared to the reported 9.1× action-head speedup. While it helps meet the 15 Hz control deadline in the authors' setup, I am not sure how much additional benefit it would provide in practical deployments. In particular, techniques such as real-time chunking (RTC) can already overlap inference with action execution and hide much of the inference latency. I would like to see a comparison with RTC to understand whether MAWS offers a meaningful advantage in real-time control.

Questions for authors

Please see weaknesses

6 comments
AuthorsOpenReview ↗

Response to Reviewer 9x4s (1/4)

We thank the reviewer for the careful reading and address the three concerns in turn, each with new experiments. Numbers taken from the paper are marked as such.

W1. Novelty: what MAWS contributes beyond chunk reuse

The novelty is not that the previous chunk is reused, which the paper credits to prior work (Sec. 2.3), but what turns that chunk into a valid starting point for a single step. MAWS contributes a diagnosis of when and why one step fails in closed loop, and the marginal constraint under which one step from the previous chunk suffices. The prior-chunk methods do not use that constraint, and without it the same chunk is the worst option in Table 2.

Diagnosis. One step from noise fails on some tasks and not on others. The split is invisible in open loop, where the one-step residual is small on every task (Sec. 3.1), and is decided in closed loop, where the residual persists and integrates in the state until it exceeds the task's tolerance (Sec. 3.2). A warm start has to lower that persistent residual, not smooth the output.

Constraint. A flow head is trained on the marginal x_s = s·a + (1−s)·ε, so at flow time s its input must be the answer plus noise of width exactly 1−s. A clean previous chunk is not on that marginal, and the prior-chunk methods pay for this in steps (Falcon needs 7.6 evaluations per call to reach 75.6 on π0 Fractal, one-step MAWS 77.4, Table 2). Re-noising has two ingredients, a prior that is a sample of the policy's own action distribution and noise of the width the marginal prescribes. Two new experiments remove one at a time.

Removing the width. We keep the prior and s fixed and scale only the noise, x = s·g + c·(1−s)·ε, on the two Fractal tasks where the warm start gains (π0, s = 0.5, one step, 3 seeds, 324 episodes per cell). If the gain came from the chunk itself, less noise would keep more of it. If it comes from meeting the marginal, success peaks at c = 1.

π0 Fractal, SR (%)c = 0.5c = 0.75c = 1 (marginal)c = 1.25one step from noise
Open Drawer54.354.659.957.454.0
Apple17.317.925.622.516.4

Success peaks at c = 1 on both tasks. Under-noised, the same prior falls back to the level of one step from noise, and with no noise at all it is the RTI-DP-style collapse of Table 2 (9.9 and 18.5). Over-noised, it gives part of the gain back. The gain comes from re-noising the chunk to the marginal, not from the chunk as such.

Removing the sample. A prior built from the current robot state, re-noised with the same s, has the right width but is not a sample of the action distribution. It never beats one step from noise (π0 Fractal 72.6 vs 73.6, ALOHA 2.0 vs 22.0) and collapses where it covers the whole chunk (CALVIN, r = H: 6.9 vs 25.1, with MAWS at 40.7). All 3 seeds.

Re-noising is a one-line operation, the shifted chunk entering the flow at s with fresh noise of width 1−s instead of entering clean, but it is not a detail. The same shifted chunk gives 9.9 without it and 77.4 with it at one evaluation per call, level with ten evaluations from noise (77.2), while the prior-chunk methods that skip it need several evaluations to come within two points (Table 2). To our knowledge none of them identifies the constraint, and it is what turns chunk reuse into a one-step sampler.

AuthorsOpenReview ↗

Response to Reviewer 9x4s (2/4)

W2. Temporal smoothing: the gain is not an averaging effect

We ran the comparison the reviewer asked for on π0 Fractal (official protocol, 3 seeds, category-weighted average). The controls give smoothing the same chunk and the same weight as MAWS, applied after the head instead of inside the step, so whatever they recover is smoothing alone:

  • TE: the temporal ensembling of ACT [1], which averages the predictions that overlapping chunks make for the same timestep. With H = 4 and r = 2 every timestep is predicted by exactly two chunks, so the executed chunk is (1−λ)·new + λ·previous over the overlapping positions, λ = 0.5 (ACT's near-uniform weight, and the retention s of MAWS).
  • TE-EMA: the same over all positions with the MAWS prior (shifted chunk, repeated tail), carried forward, i.e. an exponential moving average over chunks.
π0 Fractal, SR (%)NFECokeMoveDrawerAppleAvg
from noise (paper)1097.377.466.442.677.2
from noise (paper)197.378.264.016.473.6
from noise + TE197.167.659.322.870.2
from noise + TE-EMA193.169.058.018.568.3
+MAWS (paper)199.082.467.725.677.4

Smoothing the one-step output lowers the average below one step from noise (70.2 and 68.3 vs 73.6), costing 9–11 points on Move Near and 5–6 on Drawer, while MAWS gains on every category with the same chunk and the same weight. Three observations separate the mechanisms:

  • MAWS moves the output toward the full-step answer, not toward the prior: the prior sits 0.08–0.25 from the posterior mean, farther than the one-step output, yet the step taken from it lands closer than either (residual lowered by 17–35% on every task, Table 3). Averaging with the prior cannot move the output away from the prior.
  • The executed actions are not what carries the gain: with r = 2 of H = 4, every tail fill of Sec. 5.4 (Fig. 5) shares the same executed positions and differs only in the two unexecuted tail positions, yet zeros there return success to the from-noise level and any fill that continues the motion gains 5–9 points.
  • A smoothing effect would be monotone in how much of the prior is kept, whereas the sweep above peaks exactly at the training marginal.

Our sentence describing the loop as "the temporal ensemble of ACT inside the sampler" (Sec. 4.2) invites the reviewer's reading. The analogy was to the memory over chunks, not to the mechanism, and we will reword it.

[1] T. Z. Zhao, V. Kumar, S. Levine, C. Finn. Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware. RSS 2023.

AuthorsOpenReview ↗

Response to Reviewer 9x4s (3/4)

W3a. End-to-end latency: the head's share of the call sets the ceiling

The reviewer is right that 67.9 → 54.7 ms is modest next to the 9.1× head speedup. A VLA call has two stages, the VLM backbone and the action head, and MAWS acts on the head, so its end-to-end gain is capped at call / backbone. On the real-robot π0.5 (RTX 5090) the backbone is 39.7 of the 67.9 ms, so no head-side method can exceed 1.35×. MAWS removes 90% of the head (17.6 → 1.8 ms) and reaches 1.24×. The cap binds backbone acceleration just as hard: on π0 the head is 82% of the call, so an infinitely fast backbone would give at most 1.22×.

Model (default steps from noise)head sharehead: default → MAWS, 1 step (ms)E2E: default → MAWS, 1 step (ms)E2E speedup
π0 (10)82%322.5 → 33.8394.5 → 108.73.6×
Fast-WAM (10)83%516 → 54622 → 1643.8×
VLANeXt-B (10)79%372 → 43470 → 1423.3×
GR00T N1.7 (4)62%91.9 → 25.1147.4 → 80.21.8×
π0.5 (10)31%37.0 → 6.0120.8 → 89.81.3×
π0.5, real robot, RTX 5090 (10)26%17.6 → 1.867.9 → 54.71.2×

Batch 1, isolated, A40 unless noted. Fast-WAM (a world action model, Wan2.2-5B video DiT with a flow-matching action expert) and VLANeXt-B (Qwen3-VL-2B with a flow head) were added for this rebuttal, with released checkpoints and the paper's rule for s. The end-to-end gain tracks the head's share: 3.3–3.8× where the head dominates, 1.2–1.3× where the backbone does. The real-robot number is the smallest in the paper because π0.5 has the smallest head share, not because MAWS does less there. The two sides compose. We combined MAWS with visual-token pruning of the backbone, SAFE-Pruner [2], on π0.5 / LIBERO (4-suite average, 500 episodes per suite, pruned rows including the periodic full-token keyframes):

backboneheadSR (%)call (ms)speedup
densefrom noise, 10 steps95.21311.00×
prunedfrom noise, 10 steps94.81131.16×
denseMAWS, 1 step97.8851.54×
prunedMAWS, 1 step96.3711.84×

Either side alone stalls on the other stage, together they nearly halve the call with success above the dense baseline. In synchronous control the robot idles for the whole call at every replan, so the saving reaches the task: the same trajectory finishes about 1.37× sooner on π0 Fractal, and the completion time of successful Fast-WAM LIBERO episodes drops from 11.4 to 6.6 s (Spatial) and 27.0 to 15.9 s (Long).

[2] S. Ma, C. Zhang, C. Wang, Y. Wang, Y. Wu, Z. Wang, J. Tian, Z. Zhu, Y. Tang. SAFE-Pruner: Semantic Attention-Guided Future-Aware Token Pruning for Efficient Vision-Language-Action Manipulation. ECCV 2026. arXiv:2605.29662.

Reviewer_euyFOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents Marginal-Aligned Warm Start (MAWS), a training-free approach to accelerating flow-based VLA action heads through single-step inference. The authors investigate why single-step action generation succeeds on some tasks but fails on others, attributing the failures to persistent prediction errors accumulated during closed-loop execution. Based on this observation, MAWS reuses previously predicted action chunks with marginal-aligned re-noising to improve single-step predictions. Extensive experiments across multiple VLA models, simulation benchmarks, and real-world tasks demonstrate promising improvements in inference efficiency while largely preserving task performance.

Critical strengths
  1. Interesting Observations and Analysis. The paper provides an interesting analysis of single-step action generation, highlighting the importance of persistent prediction errors and task-dependent error tolerance in closed-loop execution.

  2. Simple and Effective Method. MAWS is a simple, training-free approach that can be integrated into existing flow-based VLA models without modifying model parameters or requiring additional training.

  3. Extensive Experimental Validation. The authors evaluate the proposed method across multiple VLA models, simulation benchmarks, and real-world robotic tasks. The results demonstrate promising efficiency improvements, achieving up to 9.1× acceleration of the action head while maintaining competitive task performance.

Critical weaknesses

Generalization to More Architectures and Challenging Benchmarks. Given the simplicity and training-free nature of MAWS, broader experimental validation would further strengthen the paper. In particular, it would be valuable to evaluate the method on more challenging benchmarks, such as RoboDojo and LIBERO-Plus. Furthermore, applying MAWS to additional flow-based VLA models, such as VLANeXt, and recent World Action Models (WAMs), such as Fast-WAM, would provide stronger evidence of its generality and plug-and-play applicability.

Questions for authors

See weakness.

4 comments
AuthorsOpenReview ↗

Response to Reviewer euyF (1/2)

We thank the reviewer for the positive assessment and the concrete suggestions. We applied MAWS, unchanged, to the two architectures the reviewer named and evaluated it on LIBERO-plus. All numbers below are new.

Overview

New modelArchitectureBenchmarksRetention s
VLANeXt-BQwen3-VL-2B + flow-matching policy head, chunk of 8, all executed (r = H)LIBERO (4 suites), LIBERO-plus (7 perturbation types)0.25 (the paper's r = H rule)
Fast-WAMworld action model: Wan2.2-5B video DiT + flow-matching action expert, chunk of 32 with 10 executed (r < H)LIBERO (Spatial, Long)0.5 (the paper's r < H rule)

Nothing was tuned: s follows the paper's rule, the first call of an episode is one step from noise, and everything else is the model's own sampler.


1. VLANeXt-B on LIBERO. Released per-suite checkpoints and protocol, 10 tasks × 50 trials = 500 episodes per suite, seed 42. The checkpoint defaults to 10 denoising steps, the released evaluation config uses 5. Latency is isolated on one A40 at batch 1.

VLANeXt-BNFEbackbone (ms)head (ms)E2E (ms)SpatialObjectGoalLongAvg (%)
from noise1010037247097.899.096.693.896.8
from noise (released setting)59919129198.299.696.494.497.2
from noise11014514697.898.895.495.096.8
+MAWS1984314297.898.696.495.697.1

This model is already step-tolerant on LIBERO: one step from noise loses nothing on average, the "often enough" case of our title. MAWS matches the released 5-step setting (97.1 vs 97.2) at one fifth of the head evaluations and half the call latency (142 vs 291 ms). Because the benchmark is near its ceiling, we turned to LIBERO-plus for a setting in which one step does lose.

2. VLANeXt-B on LIBERO-plus. Of the two harder benchmarks the reviewer named we chose LIBERO-plus. RoboDojo's policies are fine-tuned on its own data and the best of them reaches 8.8% success in simulation (π0.5 6.9%), too close to the floor to measure whether a sampler preserves success, whereas on LIBERO-plus the same checkpoint scores between 57% and 94% depending on the perturbation. LIBERO-plus applies 7 perturbation types over the 4 suites, 1 trial per task, 10,030 tasks per variant. Within the discussion period we evaluated a stratified random 25% subset (seed 0, stratified by perturbation type), the same 2,519 tasks for every variant, so the comparisons are paired.

VLANeXt-B, SR (%)NFESpatial (603)Object (632)Goal (651)Long (633)Avg (all 2,519 tasks)
from noise (released setting)593.283.572.278.281.6
from noise191.981.570.276.979.9
+MAWS193.582.470.477.780.8

Under perturbations one step from noise does lose: 1.7 points below the released 5-step setting over all 2,519 tasks (2,013 vs 2,055 successes), and below on every suite. MAWS recovers about half of that loss (2,035 successes, +0.9 points) and ends 0.8 points below the 5-step setting at one fifth of the head evaluations. The recovery is partial, and we state it as such. By perturbation type, the one-step loss is concentrated in background, language and layout. MAWS recovers it fully for background, partly for layout and not for language, while camera, light and sensor noise are flat for all three samplers.

AuthorsOpenReview ↗

Response to Reviewer euyF (2/2)

3. Fast-WAM. A world action model, a different model class from the VLAs in the paper: a Wan2.2-5B video diffusion transformer with a flow-matching action expert, released LIBERO checkpoint, first-frame mode, 10 denoising steps by default, a chunk of 32 actions of which 10 are executed. The video transformer's KV cache is pre-filled once per call and the action expert runs its denoising loop against it. MAWS applies to that loop unchanged. Protocol: Spatial and Long, 10 tasks × 10 trials (100 episodes per suite, reduced for the discussion period). Latency is isolated on one A40 at batch 1.

Fast-WAMNFEbackbone (ms)head (ms)E2E (ms)Spatial (%)Long (%)
from noise (default)1010751662296.093.0
from noise11065315897.094.0
+MAWS11105416499.097.0

MAWS has the highest success on both suites and cuts the head 9.5× and the call 3.8× (622 → 164 ms). The head is 83% of the call on this model, so the head-side speedup carries through to E2E.

Summary. MAWS has now been applied unchanged to five flow-based policies of two kinds (four VLAs with a flow head, π0, π0.5, GR00T N1.7 and VLANeXt-B, and one world action model, Fast-WAM) on SimplerEnv, LIBERO, LIBERO-plus, CALVIN, ALOHA and a real robot, always with the paper's rule (s = 0.5 for r < H, 0.25 for r = H, first call from noise). Where one step from noise already suffices (VLANeXt-B on LIBERO), MAWS costs nothing. Where it loses (LIBERO-plus, Fast-WAM Long), MAWS recovers part or all of the loss at one evaluation per call. We are happy to provide further details.

Thanks for the reply. Most of my concerns are solved, I will raise my score to positive.

We thank the reviewer for reading our response and for raising the score. The suggested architectures and benchmark led to results that strengthened the generalization evidence, and we are grateful for them. We remain available for any further question before the discussion closes.

Reviewer_9hRdOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes MAWS, a training-free method for single-step inference in flow-matching VLA action heads. It shifts the previous action chunk, re-noises it, and performs one Euler update. Experiments across simulation benchmarks and real robots show substantial speedups and improvements over single-step sampling from noise, although performance relative to full-step inference varies across tasks.

Critical strengths
  1. The method is simple and inexpensive, requiring neither retraining nor changes to model weights, which makes it attractive for deployment.
  2. The evaluation spans multiple models and environments, and the analysis of persistent errors provides a useful perspective on single-step inference in closed-loop control.
Critical weaknesses
  1. Results are mixed across tasks, and the practical advantage over moderate sampling budgets, such as four steps, remains insufficiently established. Action-head speedups alone do not fully capture deployment benefits; broader comparisons of success rate and end-to-end latency are needed.
  2. The necessity of using the previous predicted trajectory is unclear. A comparison against initialization from the current measured robot state would help establish whether prediction history provides benefits beyond a simpler state-based prior.
Questions for authors
  1. Could you provide broader comparisons between MAWS and fixed four-step inference using the same models and benchmarks, reporting success rate, action-head latency, and end-to-end latency? Beyond the existing GR00T comparison, results on tasks with performance degradation would help assess whether the additional speedup justifies the loss.
  2. Have you tried constructing the initial prior from the current robot state, appropriately mapped into the action representation, and applying the same re-noising and single-step update? This would test correction from a state-based starting point, particularly when r = H. Does the zero-fill ablation already cover this case under your action representation and normalization?
4 comments
AuthorsOpenReview ↗

Response to Reviewer 9hRd (1/2)

We thank the reviewer for the constructive questions. We ran every requested experiment. All numbers below are new.


W1 / Q1. Fixed four-step inference on the same models and benchmarks

We added four steps from noise for π0 on both SimplerEnv suites (official protocols, 3 seeds, mean ± s.d.), complementing the GR00T comparison. Latency is isolated on one A40 at batch 1, E2E = VLM backbone + action head.

π0 (SimplerEnv)NFEhead (ms)E2E (ms)Fractal SR (%)Bridge SR (%)
from noise (default)10322.5394.577.2 ± 0.868.2 ± 0.8
from noise4127.9199.178.6 ± 1.471.5 ± 0.3
from noise132.9104.273.6 ± 1.171.3 ± 0.9
+MAWS133.8108.777.4 ± 0.974.7 ± 1.1
+MAWS263.8134.879.1 ± 1.273.4 ± 0.5
  • On Bridge, one-step MAWS beats four steps from noise by 3.2 points at a quarter of the head latency (34 vs 128 ms).
  • On Fractal, one-step MAWS is within seed variation of four steps from noise (77.4 ± 0.9 vs 78.6 ± 1.4) at 3.8× lower head and 1.8× lower E2E latency. The gap is entirely Apple, the longest-horizon task: on Coke, Move and Drawer one-step MAWS matches or exceeds four steps from noise (99.0 / 82.4 / 67.7 vs 99.1 / 81.1 / 66.6), on Apple it scores 25.6 against 40.1. Apple is the collapse case of the paper's analysis (Sec. 5.4, Table 3): MAWS lowers the one-step residual there from .047 to .034, but the task's tolerance is below .05, so a single step from any starting point stays above it and the remedy is a second step, not a different prior. In the paper's step sweep (appendix), MAWS reaches 41.4 on Apple at two steps and 47.5 at three, against 18.2 and 31.2 from noise at the same budgets and 42.6 at the checkpoint's ten steps.
  • Two-step MAWS matches or exceeds four steps from noise on both suites (79.1 vs 78.6, 73.4 vs 71.5) at half the head latency (64 vs 128 ms) and 1.5× lower E2E latency.

The step count is a deployment choice. One-step MAWS gives the largest speedup where one step suffices, and where more accuracy is needed, warm-started steps buy it more cheaply than steps from noise.


W1. End-to-end deployment benefit

We agree that E2E latency is what matters in deployment, and the arithmetic behind it is worth making explicit. A VLA call has two stages, the VLM backbone and the action head, and an acceleration method acts on one of them. MAWS acts on the head. Accelerating one stage leaves the other as the bottleneck, so the E2E gain of a head-side method is capped at call / backbone and that of a backbone-side method at call / head. The cap binds backbone acceleration just as hard.

  • On π0 the head is 82% of the call: an infinitely fast backbone would cut E2E by at most 1.22×.
  • On our real-robot π0.5 the backbone dominates: an infinitely fast head would cut E2E by at most 1.35×.

The two sides are therefore complementary. Across the models we measured, MAWS alone gives 3.3–3.8× E2E where the head dominates (π0, Fast-WAM, VLANeXt-B) and 1.2–1.3× where the backbone does (π0.5 in simulation and on the real robot). Combining the two sides gives far more than either alone, as the next experiment shows.

AuthorsOpenReview ↗

Response to Reviewer 9hRd (2/2)

Combining backbone and head acceleration (W1, continued). We built one such combination. For the backbone we chose a visual-token pruning method, SAFE-Pruner [1], which prunes image tokens inside the VLM pass and leaves the action head untouched. For the head we used MAWS. The two act on different stages and neither needs the other: the pruned backbone's output feeds the MAWS step. Setting: π0.5, LIBERO, 4-suite average, 500 episodes per suite. Latency is isolated on one A40 at batch 1, and the pruned rows are the effective average including the periodic full-token keyframes (about 25% of calls in the closed-loop runs).

backboneheadSR (%)backbone (ms)head (ms)call (ms)speedup
densefrom noise, 10 steps95.277511311.00×
prunedfrom noise, 10 steps94.861471131.16×
denseMAWS, 1 step97.8766851.54×
prunedMAWS, 1 step96.3616711.84×

Each method alone stalls on the other stage: pruning leaves 47 ms of head (1.16×), MAWS leaves 76 ms of backbone (1.54×). Together they nearly halve the call (1.84×) with success above the dense 10-step baseline (96.3 vs 95.2). Pruning alone costs 0.4 points of success (94.8), as token pruning usually does, MAWS alone raises it by 2.6 points (97.8), and the combination keeps most of that gain while taking both speedups. MAWS is a training-free, drop-in head-side component that composes with any backbone accelerator.


W2 / Q2. Is the previous chunk necessary? A state-based prior

We replaced the MAWS prior with one built from the current robot state, keeping everything else identical: the same s, re-noising, single step and first call from noise. The state prior is the chunk that would keep the robot where it is, in the model's normalized action space and repeated over all H positions, so it has the shape and scale of a predicted chunk but carries nothing from the previous call:

  • Delta end-effector actions (SimplerEnv, CALVIN): the six motion dimensions hold zero displacement in the checkpoint's normalized action units (not the zero vector), and the gripper dimension holds the command that keeps the gripper at its current opening, read from the proprioceptive state.
  • Absolute joint actions (ALOHA): the current joint positions from the observation, re-normalized from state to action statistics, repeated over the 50-step chunk.
π0executed / horizon (r / H)from noise, 1 stepstate prior, 1 stepMAWS, 1 stepfrom noise, default steps
SimplerEnv Fractal (%)2 / 473.672.677.477.2 (10 steps)
ALOHA TransferCube (%)10 / 5022.02.032.734.0 (10 steps)
CALVIN ABC→D, chain of 5 (%)5 / 525.16.940.728.0 (5 steps)

All rows: 3 seeds (Fractal category-weighted average, ALOHA 100 episodes per seed, CALVIN 1000 chains per seed).

The state prior never beats one step from noise. On Fractal it is within a point of it (72.6 vs 73.6), on ALOHA it collapses (2.0 vs 22.0), and on CALVIN, the r = H case the reviewer highlighted, it collapses to 6.9 while MAWS reaches 40.7, 15.6 points above one step from noise. We attribute the difference to how the head reads its starting point. At flow time s it was trained to see xsx_s as the answer plus noise of width 1−s1−s, so a single step should move toward whatever xsx_s is centered on. The previous chunk is the policy's own prediction for the same timesteps, so it is centered near the answer and, in our reading, carries the motion underway, which a memoryless policy cannot recover from the current observation alone. The current state is centered on standing still, a confident wrong answer in the middle of a task, so the step would be pulled toward no motion. The table is consistent with this reading: across the three benchmarks, the gain of the right prior and the loss of the wrong one grow together.

We hope these results resolve the reviewer's concerns, and we are happy to provide further details.

[1] S. Ma, C. Zhang, C. Wang, Y. Wang, Y. Wu, Z. Wang, J. Tian, Z. Zhu, Y. Tang. SAFE-Pruner: Semantic Attention-Guided Future-Aware Token Pruning for Efficient Vision-Language-Action Manipulation. ECCV 2026. arXiv:2605.29662.

Thank you for your response, and it has cleared up some of my questions. I will keep my score as it is. Best regards.

We thank the reviewer for the positive assessment and for confirming that our response cleared up some of the questions. The reviewer's requests led to results that sharpened the paper, and we are grateful for them. We remain available for any remaining question before the discussion closes.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.