acceptodds
Under review as a conference paper at ICLR 2027

SnapFlow: One-Step Action Generation for Flow-Matching VLAs via Progressive Self-Distillation

23 views

Abstract

Vision-language-action (VLA) models with flow-matching action heads, such as 0 and 0.5, have become a standard recipe for generalist robot manipulation, but every action chunk is produced by integrating a learned velocity field over about ten sequential action-expert evaluations, so action generation dominates model latency and caps the achievable control frequency. Existing remedies either shrink each evaluation through layer distillation or pruning while leaving the sequential solver in place, or import one-step objectives from image generation that were designed for training from scratch, rely on an exponential-moving-average teacher or an external critic, and regress finite-span targets onto the conditional velocity, whose variance around the marginal field biases the learned map. We introduce SnapFlow, a self-distillation method that adapts a pretrained flow-matching VLA to one-step action generation. SnapFlow mixes local flow-matching supervision with shortcut samples whose target is the model's own two-step Euler endpoint under stop-gradient, and adds a single zero-initialized destination-time projection so that one action expert serves both objectives while the pretrained predictor is preserved at initialization; no separate teacher network or critic is required. On 0.5, whose 10-NFE teacher operates near ceiling, the archived one-NFE SnapFlow checkpoint reaches 98.75% LIBERO success versus 97.75% for the teacher; under a separate paired evaluation of 0, the one-NFE model surpasses its teacher by 10.5 percentage points. SnapFlow also lowers mean and tail action error on held-out data, cuts batch-one model latency by about 3.3, and transfers to shelf retrieval on a dual-arm humanoid.

Primary area:
applications to robotics, autonomy, planning
Keywords:
flow matching, vision-language-action models, VLA acceleration, self-distillation, one-step generation, robot manipulation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_XUEUOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces SnapFlow, a self-distillation method for VLA models. Its core idea is to combine standard flow-matching objective with two-step Euler endpoints objective generated by the model itself, alone with a zero-initialized destination-time projection. In simulation experiments, the method successfully distills pi0 and pi05 from 10-step to one-step action generation, achieving more than 3 times inference speedup without sacrificing task success rates. The paper also presents several toy real-robot deployment demonstrations.

Critical strengths
  1. The proposed SnapFlow maintains high success rates on the LIBERO benchmark.
Critical weaknesses

Main:

  1. There are many ways to distilling flow-matching models and flow0matching VLAs into on-step model, as the authors say in the introduction and related work sections. However, the paper does not provide quantitative comparisons and analysis with these methods, leaving many of its claims insufficiently supported.
  2. The experimental gains appear minor. In simulation, naive single-step generation already performs very similarly to the teacher, and SnapFlow offers only small additional gains over both. Naive single-step generation also appears to perform reasonably well. The real-robot experiments provide only a few toy video demonstrations, without quantitative results. It seems the paper should evaluate the method on more challenging simulation benchmarks to better demonstrate its effectiveness.
  3. The performance of 2-, 3-, 4-, and 5-step generation is reported only in terms of offline action quality, without corresponding online success rates. To my knowledge, many real-robot and simulation evaluations use 4 or 5 denoising steps by default to balance inference speed and task performance [1][2][3]. The authors should establish whether simply SFT on the official LIBERO dataset with a 4-step configuration could achieve comparable or even better performance.
  4. It will be better to report end-to-end throughput or task completion speed in both simulation and real-world experiments. Robotic tasks involve more than model inference. The time spent interacting with the environment also matters. Based on a rough visual assessment, SnapFlow may improve actual throughput by only 30–40% over the teacher, which is substantially less pronounced than the reported 3 times model inference speedup.

Minor:

  1. The legend in Figure 2(a) contains four items, but only two of them are visible in the plot.

[1] Chen, Kang, et al. "πrl: Online rl fine-tuning for flow-based vision-language-action models." arXiv preprint arXiv:2510.25889 8 (2025). [2] Black, Kevin, Manuel Galliker, and Sergey Levine. "Real-time execution of action chunking flow policies." Advances in Neural Information Processing Systems 38 (2026): 33383-33407. [3] Wang, Chenghua, et al. "PhyAI: Real-Time Physical AI at the Edge, Scalable Rollouts in the Cloud." arXiv preprint arXiv:2608.03682 (2026).

Questions for authors
  1. The experimental results refer only to LIBERO, but LIBERO comprises multiple benchmark suites, such as LIBERO-goal. Which suites, tasks, or combinations of tasks were used in the evaluation?
  2. How does the distillation data organized? How sensitive is SnapFlow to the distribution, quality, and quantity of the distillation data?
4 comments
AuthorsOpenReview ↗

Response to Reviewer XUEU (1/2)

We first answer the two questions, then the weaknesses. The general response gives the full paired tables.

Q1. Which LIBERO suites and tasks were used?

All four standard suites, LIBERO-Spatial, LIBERO-Object, LIBERO-Goal and LIBERO-Long (libero_10), each with its 10 tasks and 10 episodes per task (400 episodes per method and evaluation seed). Table 1 reports the per-suite and average success rates (columns Spatial, Object, Goal and Long-10, the last being libero_10, as the caption now states). Appendix C lists every one of the 40 tasks for each method in the paired π0.5 runs, and Appendix D.2 gives the π0 counts for every evaluation seed and suite.

Q2. How is the distillation data organized, and how sensitive is SnapFlow to it?

SnapFlow uses no extra data. It fine-tunes on the same LIBERO demonstration set that the teacher was fine-tuned on (the LeRobot release of LIBERO, all four suites, 1,693 episodes, 273k frames). In the revised recipe every training example is one observation, its instruction and a noised version of the demonstrated action chunk; the target is the endpoint that two Euler steps of the frozen pretrained model reach from that noised input, so the demonstrated action enters the loss only through the noised input and no action label is regressed directly (the submitted recipe mixed such rows with ordinary flow-matching rows at α = 0.5). Training runs 30k micro-batches of 4, accumulated over 8 into 3,750 updates of effective batch 32, i.e. 120k examples or about 0.44 passes over the data. We have not measured sensitivity to the quantity or quality of the distillation data and say so in Section 4.1.

W1. No quantitative comparison with other distillation methods.

Same checkpoint, data and budget, paired closed-loop evaluation, one-step inference, public π0 LIBERO checkpoint, 3 evaluation seeds (n = 1200): naive one-step without any training 1127/1200 (93.9%); flow-matching fine-tune with the same budget 1101 (91.8%); submitted SnapFlow recipe after two implementation fixes 1084 (90.3%); submitted recipe with frozen-teacher target 1079 (89.9%); revised SnapFlow recipe, as defined in the general response and not the method as submitted, 1125 (93.75%); 10-step teacher 1134 (94.5%). On π0.5 (aggregate): frozen-teacher target 97.25, EMA target 98.00, 10-step online target 97.00, naive one-step 98.25. The revised recipe is the only fine-tuned student indistinguishable from the teacher (p = 0.41), and no method, ours included, is distinguishable from naive one-step inference (p = 0.92 for the revised recipe). A Shortcut Models baseline with unrestricted step sizes was not completed in time and is not reported. On LIBERO these comparisons cannot separate any one-step method from naive one-step inference, because the teacher itself gains nothing from its extra steps; the revision states this plainly.

W2. The gains are minor; harder benchmarks.

On LIBERO the gain over the teacher is zero, and the revision says so. The paired 10-step-versus-1-step difference of the teacher itself, which measures how much a one-step student could recover from the teacher's sampling, is −0.6 pp on π0 (p = 0.55, n = 1200) and 0.0 pp on π0.5 (n = 400). On LIBERO-Plus, a fixed paired set of 400 perturbed variants across five perturbation dimensions, the teacher, its naive one-step sample and the revised one-step student all succeed on 263/400 (65.8%; p = 1.0 for every pair; Appendix D.3), so distribution shift does not open a gap either. On PushT (lerobot/pusht, the LIBERO π0 weights fine-tuned for 30k steps; 200 paired episodes, one training and one evaluation seed; Appendix D.4), where the demonstrations are multimodal, the same teacher loses 17.5 pp at one step (30.0% vs 12.5%, p < 0.001). The submitted recipe with frozen-teacher target (batch 8) recovers 13.0 pp of it (25.5%; p = 0.001 vs naive one-step; −4.5 pp vs the teacher, p = 0.35) and the revised recipe 11.5 pp (24.0%; p = 0.002 vs naive; −6.0 pp vs the teacher, p = 0.22), but a control that only continues flow-matching training recovers 8.5 pp (21.0%), and the three are statistically indistinguishable (p = 0.31 and 0.55 against the control); the revised recipe also used half the samples and an eighth of the optimizer steps of the control, so its number is not an equal-budget comparison. We therefore reposition the contribution as recovering 10-step accuracy at one step with a frozen VLM and no deployment change, state explicitly that on LIBERO the teacher's own one-step sample already does this, and claim no advantage over continued flow-matching training where the headroom is positive. The real-robot material is presented as a qualitative check only.

AuthorsOpenReview ↗

Response to Reviewer XUEU (2/2)

W3. Closed-loop results at 2 to 5 steps; would 4-step SFT suffice?

Teacher at 1/2/3/4/5/10 NFE, paired, seed 1000, 400 episodes (the 2-NFE run lacks LIBERO-Spatial): 377, 278 of 300, 374, 385, 379 and 373 successes. Four steps is the best single-seed number and the other step counts are within 12 episodes of it, inside the ±4 pp single-seed CI, and over three seeds 10 steps and 1 step differ by −0.6 pp. Flow-matching SFT does not depend on the inference step count, so "SFT with a 4-step configuration" is the existing fine-tuned teacher run at 4 NFE, which is in the table. On LIBERO it performs the same as 1 NFE at the higher latency reported in Table 8 of the submission (Appendix H of the revision), so 4-step inference is a reasonable default there but not a better one. On PushT the teacher reaches 12.5%, 22.5% and 30.0% at 1, 2 and 10 NFE (n = 200, paired; 2 vs 10 steps −7.5 pp, p = 0.09), so there the step count does matter; 3 to 5 steps were not run on PushT. The teacher sweep is in Appendix O of the revision and the PushT numbers in Appendix D.4.

W4. End-to-end throughput and task-completion time.

Agreed: the 3.3× figure is model-inference latency only, and the episode-level effect is much smaller because simulation and environment stepping dominate. We did not complete a per-episode wall-clock measurement within the discussion period, so the revision makes no claim about throughput or task-completion time: it reports 3.3× as batch-one model latency, adds that naive one-step decoding gives the same speedup, so the case for distillation cannot rest on speed, and states that no real-robot timing was measured under a controlled protocol (Sections 4.5 and 5). The supplementary side-by-side videos use a timer derived from policy steps × the measured per-chunk latency, as the supplementary README states, not wall-clock time; they therefore overstate the episode-level speedup, and we no longer cite them as evidence of it.

Minor. Figure 2(a) legend.

Thank you. Two of the four series were hidden: the SmolVLA markers coincided with the π0.5 markers after per-model normalization, and the π0.5 SnapFlow marker lay below the axis range. The revised Figure 2(a) offsets the SmolVLA markers and extends the axis so that all four series are visible; panel (b) now shows the paired numbers of Table 1.

Summary of changes for this review: paired multi-seed results replace the unpaired numbers; same-budget flow-matching control and teacher at 2 to 5 NFE added; LIBERO-Plus and PushT added, with a headroom table for every base policy and benchmark; the suites, the distillation data and the budget are stated explicitly; the episode-time claim is removed; the real-robot material is qualitative only; Figure 2(a) is fixed.

Reviewer_XUEUOpenReview ↗

Response to authors

Thank you for your reply and clarification. Some issues (such as the detailed settings of the LIBERO evaluation, the configuration of distillation data, etc.) have been resolved. I still have the following two main concerns:

(1) Many numerical results in the paper remain unconvincing due to the issue of teacher's NFE. Given that NFE=4 is already a reasonable default setting for the teacher, the reported 3.3x speedup is still calculated relative to a 10-step teacher. This result lacks justification. Should the paper instead report the speedup relative to an NFE=4 teacher?

(2) The experiments in the paper are still not sufficiently convincing. For instance, comparisons with several sota one-step flow/diffusion models (such as Shortcut Model, MeanFlow, Consistency Policy, etc.) are missing. Furthermore, while it is reasonable for less performance improvement over the teacher as a VLA acceleration work, the improvement compared to a naive 1-step approach remains marginal. Additionally, since the LIBERO benchmark itself is relatively simple and success rates across all methods are quite close, SnapFlow has not been evaluated on more challenging benchmarks. Therefore, I still have concerns about the real effectiveness of SnapFlow.

AuthorsOpenReview ↗

Response to Reviewer XUEU (3/2)

Thank you for the follow-up, and for confirming that the questions on the evaluation suites and the distillation data are resolved.

(1) Speedup relative to a 4-step teacher. We agree that 10 NFE is not the only reasonable reference, and that 3.3× overstates the gain against a teacher run at a practical step count. Table 13 of the revision (A800, batch 1) times the pretrained π0.5 solver at 1, 2, 3, 5 and 10 NFE: 81.0, 103.3, 124.4, 166.9 and 274.0 ms. The one-step student runs at 83 ms, so against the teacher at 3 and 5 NFE the speedup is 1.5× and 2.0×. We did not time 4 NFE, but each extra step costs about 21 ms, which puts a 4-step teacher near 145 ms and the speedup against it at about 1.75×. The VLM prefix (about 60 ms) is paid at every step count, which is why the speedup is far below the reduction in action-expert calls. On LIBERO the 4-step teacher succeeds on 385/400 episodes against 377/400 at one step (one evaluation seed; the 95% interval of a paired difference at this size is about ±4 pp), so one step costs nothing measurable there relative to four. We will report the speedup against every timed step count, give the 4-step comparison in the main text, and no longer quote 3.3× without stating its reference.

(2) One-step baselines and harder benchmarks. We have not run Shortcut Models, MeanFlow or Consistency Policy on the same backbone, they cannot be completed before the discussion closes, and we make no claim of an advantage over them.

On harder settings, the revision goes beyond standard LIBERO (Appendices D.3 and D.4). On LIBERO-Plus (400 paired perturbed variants over five perturbation dimensions) the teacher, its naïve one-step sample and the student all drop by about 29 points, to 263/400 each, so distribution shift alone does not separate them. PushT is the setting where multi-step sampling matters: the same teacher loses 17.5 pp when sampled in one step instead of ten (30.0% vs 12.5%, p < 0.001). There the gain over naïve one-step is not marginal: the two one-step students reach 25.5% and 24.0%, +13.0 pp and +11.5 pp over naïve one-step (p = 0.001 and 0.002, 200 paired episodes), and neither is significantly below the 10-step teacher (p = 0.35 and 0.22). As we said in the general response, continued flow-matching training also recovers part of the gap (21.0%), and at this sample size the students cannot be separated from it. An equal-budget comparison with that control and with the one-step baselines you list is what would settle the question, and we agree it is the main open point.

Thank you again for reading the revision closely and for the specific suggestions.

Reviewer_Lad6OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SnapFlow, a fine-tuning method to turn pretrained flow-matching VLAs (π0\pi_0, π0.5\pi_{0.5}) into one-step policies. During fine-tuning, the VLM stays frozen and the action expert is trained on a mix of standard flow matching (s=ts = t) and one-step shortcut targets (s=0s = 0), where the target is the average velocity of a two-step Euler rollout from the current model under stop-gradient. A zero-initialized projection layer conditions the expert on the destination time ss. On LIBERO, 1-step SnapFlow on π0.5\pi_{0.5} reaches 98.75% success (vs. 97.75% for the 10-step teacher and 96.75% for naive 1-step) with a 3.3$\times$ model speedup. The paper also reports a π0\pi_0 run, offline action MSE, and a real-robot demo.

Critical strengths
  • The training recipe is simple and lightweight: it keeps the VLM frozen, needs no EMA teacher or critic, and finishes in about 12 hours on a single A800 GPU.
  • The authors are upfront about the limitations of their experiments, including the weak π0\pi_0 teacher, protocol differences, and the qualitative scope of the real-robot demo.
Critical weaknesses
  • Limited novelty and gap between theory and method. The training objective is essentially Shortcut Models [1] restricted to two step sizes (s=ts = t and s=0s = 0), combined with the destination-time conditioning from Decoupled MeanFlow [2]. Meanwhile, the theory in Sec. 3.3 mostly restates FlowConsist [3] with context cc added and does not connect to the actual implementation: Eq. 8 derives a JVP-based target, but Eq. 11 drops it for a plain two-step Euler average without analysis. Remark 2 even argues against composing local predictions, yet the two-step target does exactly that.
  • The benefit of distillation is not clearly established. On π0.5\pi_{0.5}, naive 1-step inference without any training already gets 96.75% at 81 ms, capturing almost all the speedup within 4 of 400 episodes of the teacher. SnapFlow's gains (+4 episodes over the teacher, +8 over naive 1-step) come from an unpaired evaluation with no confidence intervals and could easily be noise. SnapFlow also gets 30k extra fine-tuning steps on the target data, without a closed-loop control giving the teacher the same fine-tuning budget (Table 3a only checks offline MSE). On π0\pi_0, the "+10.5% over teacher" result comes from a poorly converged teacher (70.5%) that is already beaten by its own naive 1-step variant (77.75%). Finally, recent work [4] shows that plain flow matching with a high-noise timestep schedule already gives strong 1-step VLA performance on LIBERO, which undercuts the need for distillation and is not discussed.
  • Missing baselines and saturated benchmark. No other one-step methods are compared on the same backbone, such as Shortcut Models [1], Decoupled MeanFlow [2], OFP [5] (which also evaluates 1-step π0.5\pi_{0.5}), or MeanFlow VLA [6]. Also, all closed-loop results are on standard LIBERO, where the π0.5\pi_{0.5} teacher is already near ceiling at 97.75%. A few episodes cannot separate methods there, so harder benchmarks (e.g., LIBERO-Plus/Pro, RoboTwin 2.0) are needed.
  • Unclear practical value for deployment. The paper motivates 1-step generation through faster replanning and reactivity, yet in Appendix I, SnapFlow is 5 points worse than the teacher at nact=1n_{\text{act}} = 1 (72% vs. 77%), where fast replanning matters most. Latency is only measured on a server-grade A800 GPU with no edge-hardware or on-robot control rates, and the real-robot test (9 episodes total using separate single-target models per shelf) is too small to support any advantage claim. Lastly, the offline MSE gains in Table 2 are compared against the 10-step teacher, its worst setting in Table 9, and lower MSE may just reflect regression toward the mean action rather than better policy behavior.

References

[1] Frans et al. - One Step Diffusion via Shortcut Models - ICLR 2025

[2] Lee et al. - Decoupled MeanFlow: Turning Flow Models into Flow Maps for Accelerated Sampling - ICLR 2026

[3] Zhang et al. - FlowConsist: Make Your Flow Consistent with Real Trajectory - arXiv:2602.06346

[4] Chen et al. - Let It Be Simple: One-Step Action Generation for Vision-Language-Action Models - arXiv:2606.05737

[5] Li et al. - One-Step Flow Policy - arXiv:2603.12480

[6] Chen et al. - Mean-Flow based One-Step Vision-Language-Action - arXiv:2603.01469

Questions for authors
  1. What are the closed-loop results (at 1 and 10 NFE) if the teacher is given the exact same extra flow-matching fine-tuning budget (30k steps, frozen VLM), and how does it compare to a high-noise timestep schedule [4]?

  2. How does SnapFlow compare directly against [1] and [2] on the same π0.5\pi_{0.5} checkpoint, and what does restricting s∈{t,0}s \in \{t, 0\} add?

  3. Can you provide paired, multi-seed π0.5\pi_{0.5} results with confidence intervals, as well as results on a harder benchmark where the teacher is not near 98%?

  4. Why does SnapFlow drop 5% below the teacher at nact=1n_{\text{act}} = 1, and how does it fit with asynchronous or real-time chunking setups that already hide inference latency?

3 comments
AuthorsOpenReview ↗

Response to Reviewer Lad6 (1/2)

Several of your points are correct as stated; the revision changes accordingly. We report every new result, including the unfavorable ones; the general response gives the full paired tables.

W1. Novelty, and the gap between theory and method.

We agree that the objective combines known components: a Shortcut-style two-half-step target restricted to destinations s ∈ {t, 0}, and destination-time conditioning as in Decoupled MeanFlow. Section 3.3 did restate FlowConsist's analysis. The contribution is narrower than the submitted text suggested: a recipe that adapts a pretrained billion-parameter flow VLA to one-step generation with a frozen VLM and a shared prefix KV cache, at the cost of two extra action-expert calls per example, together with a controlled evaluation of when this helps. The revision says this in the introduction and the contribution list, and moves Theorems 1–3 to Appendix A as background from prior work.

On Eq. 8 versus Eq. 11: you are right that the submission did not connect them. The two-step Euler average is a finite-difference surrogate for the total-derivative term in Eq. 8; its error relative to the exact average velocity is second order in the step size and depends on the curvature of the learned field. The revision adds this derivation (Appendix A) and states that the two-step target is an approximation of Eq. 8, not an exact implementation. Remark 2 is rewritten: the two-step target does compose local predictions, but only to build a training target; at deployment the student takes a single direct step, so the inference-time error accumulation of Theorem 3 does not arise there.

W2 / Q1. Is the distillation benefit established? Same-budget FM control; high-noise schedule [4].

On LIBERO, no. The same-budget control (α = 1, 30k updates, batch 4, frozen VLM, fp32 master weights), run as an ordinary flow model at 1 NFE on the public π0 checkpoint with 3 evaluation seeds (n = 1200, paired), reaches 1101/1200 (91.8%): 2.8 pp below the untouched 10-step teacher (p = 0.005) and 2.2 pp below its naive one-step sample (p = 0.026). The submitted SnapFlow recipe, after the two implementation fixes described in the general response, reaches 1084/1200 (90.3%); the difference to the control, 1.4 pp, is not significant (p = 0.18). Your reading is therefore correct: on LIBERO the distillation target adds nothing measurable over plain fine-tuning, and both lose points relative to the untouched teacher. The revised recipe described in the general response (frozen-teacher target, all gradient on the one-step branch, effective batch 32; not the method as submitted) brings the student back to 1125/1200 (93.75%), indistinguishable from the teacher (p = 0.41) and from naive one-step (p = 0.92), but not above either. On π0.5 the paired evaluation gives teacher and naive one-step 98.25% and the submitted-recipe student 96.25% and 97.50% for two training seeds, so the submitted gains were noise, as you suspected; the π0 teacher of the submission was under-converged and is replaced by the public checkpoint.

We did not run the high-noise timestep schedule of [4] within the discussion period. The revision discusses [4] in Related Work and notes that on LIBERO its one-step result, like ours, faces a headroom of zero: the teacher's own one-step sample is already at teacher level, so no training recipe can be shown to add anything there. On PushT (lerobot/pusht, the LIBERO π0 weights fine-tuned for 30k steps; 200 paired episodes, one training and one evaluation seed; Appendix D.4 of the revision), where the teacher loses 17.5 pp at one step, the FM-only control does close part of the gap: 21.0% at one step (42/200) against 12.5% for naive one-step (p = 0.036) and 30.0% for the 10-step teacher (p = 0.06). The submitted recipe with frozen-teacher target (batch 8) reaches 25.5% and the revised recipe 24.0%; neither is distinguishable from the control (+4.5 pp, p = 0.31; +3.0 pp, p = 0.55; n = 200), and the control recovers coverage at least as well (0.742 vs 0.704 and 0.719). The revised recipe also used 120k samples and 3,750 optimizer steps against 240k samples and 30k steps for the control, so that comparison is not at equal budget. On PushT, therefore, the distillation target adds nothing measurable over continued flow-matching training at this sample size.

W3 / Q2, part 1: Decoupled MeanFlow [2] and the other one-step VLA methods; same-backbone numbers follow in the next post. SnapFlow already uses the destination-time conditioning of [2]; the difference lies in the target. A MeanFlow-style target requires a Jacobian-vector product through the 300M-parameter action expert under gradient checkpointing, which we did not implement; the revision presents the two-step target as an approximation of it (Section 3.4, Appendix A). OFP [5] and MeanFlow VLA [6] are described with their protocols in Related Work; we do not compare head-to-head across protocols.

AuthorsOpenReview ↗

Response to Reviewer Lad6 (2/2)

W3 / Q2 / Q3. Same-backbone baselines; what s ∈ {t, 0} adds; paired multi-seed results; harder benchmark.

Same checkpoint, same data and budget, one-step inference. On π0 (3 seeds, n = 1200, paired): submitted recipe with frozen-teacher target 1079/1200 (−4.6 pp vs teacher, p < 0.001); submitted recipe, online target 1084/1200; the same at effective batch 32 1116/1200 (−1.5 pp, p = 0.11); revised recipe 1125/1200 (−0.75 pp, p = 0.41). On π0.5 (aggregate success, initial states only partly overlapping the main evaluation): frozen-teacher target 97.25 (mean of 3 training seeds), EMA target 98.00, 10-step online target 97.00, conditional-velocity target 97.00, against 98.25 for naive one-step. The Shortcut Models baseline with unrestricted step sizes was not completed within the discussion period, and the revision says so rather than report a partial run. Given the zero headroom on LIBERO, none of these baselines, including ours, can be separated from naive one-step inference there, so the question of what restricting s adds cannot be answered on LIBERO. On PushT the frozen-target student with s ∈ {t, 0} and the revised recipe with s = 0 only are +4.5 pp (p = 0.31) and +3.0 pp (p = 0.55) above the FM-only control and −1.5 pp apart (p = 0.81, n = 200), so PushT does not separate the two choices of s from each other or from plain fine-tuning either.

Harder benchmark. We evaluated a fixed paired set of 400 LIBERO-Plus variants (80 each for background textures, language rewrites, lighting, object layout and sensor noise, across the four suites, one episode per variant, one evaluation seed; Appendix D.3):

Dimension (n = 80 each)Teacher 10 NFENaive 1 NFERevised SnapFlow 1 NFE
Background666864
Language495550
Lighting737477
Layout666366
Sensor noise936
All (n = 400)263 (65.8%)263 (65.8%)263 (65.8%)

Every pairwise difference is non-significant (p = 1.0 overall for each pair; per dimension the smallest p is 0.07, for naive versus teacher under sensor noise, and every student comparison has p ≥ 0.27). Distribution shift lowers all three policies by about 29 pp but does not open a gap between 10-step sampling and one-step inference. The submitted bf16 recipe with only the clamp fix scores 258/400 on the same variants, and the frozen-target recipe at batch 4 scores 240/400 (−5.8 pp vs teacher, p = 0.01, driven by the layout dimension). The benchmark on which multi-step sampling does matter is PushT (previous post and general response).

W4 / Q4. Practical value.

n_act = 1 (72 vs 77): one seed on 100 libero_10 episodes, so 5 points is 5 episodes, and we claim no advantage there. While checking it we found that the n_act = 10 point of the sweep had been copied from Table 1 rather than taken from the sweep run, which recorded teacher 89 / SnapFlow 91; the revision reports the sweep's own values and labels it single-seed (Appendix J). Asynchronous or real-time chunking: complementary; RTC hides latency only if inference finishes within the execution window, and fewer expert evaluations permit shorter windows or slower hardware under the same scheme. We have not evaluated SnapFlow together with RTC and say so. Hardware: latency is measured on an A800 only; we have no consumer-GPU or embedded measurement, and the revision reports 3.3× as batch-one model latency with no episode-time claim, noting that naive one-step decoding gives the same speedup. Real robot: agreed; nine episodes from per-target models support no comparative claim, and the revision presents them as a qualitative deployment check only, removed from the contribution list. Offline MSE against the 10-step teacher: agreed. The naive one-step mean MSE (.00892, from the paired sweep) is now stated next to the table; since the two measurements use different noise draws, we note it in the caption rather than add a row. The 500 samples come from the training demonstrations, not a held-out split; the revision corrects this. Because MSE rewards mean-like predictions, the revision demotes offline MSE to a diagnostic and adds a dispersion analysis (Appendix D.5): on LIBERO the teacher's own 10-step dispersion is 5.4% of the data scale and the one-step models retain 39 to 50% of it with equally accurate means.

Summary of changes: contribution narrowed; theory moved to Appendix A with the Eq. 8 to two-step-target relation and a corrected Remark 2; paired multi-seed results; same-budget FM control on LIBERO and PushT; LIBERO-Plus and PushT added; [4]–[6] discussed; horizon sweep and "held-out" wording corrected; Shortcut baseline and high-noise schedule reported as not run.

Reviewer_Lad6OpenReview ↗

Official comment by Reviewer

Thank you for the detailed response and additional experiments. The new results are informative, but they reinforce my main concerns rather than resolve them. (i) The main claim remains unsupported. Under the paired protocol, the gains over the teacher disappear, while naive one-step decoding or continued flow-matching training matches or outperforms the distillation target in every setting. This suggests that the speedup comes from one-step decoding itself, rather than the proposed training. (ii) The revised recipe differs substantially from the original method. It relies on a frozen teacher and removes the flow-matching branch used to motivate the offline ablation. Even with this revised recipe, it does not outperform the teacher. (iii) The rebuttal reveals a copied point in the horizon sweep and offline metrics computed on training data. The corrected sweep also shows substantial variation across runs of the same policy. I appreciate the authors' transparency, but these issues make the remaining single-run results difficult to assess reliably. So, I decide to keep my score.

Reviewer_KUScOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper takes a pretrained flow-matching VLA (π0.5 / π0) and fine-tunes the action expert so it can spit out an action chunk in one step instead of ten. The trick is to mix regular flow-matching training with a self-generated target (the model's own two-step Euler result), plus a small zero-init embedding so the network knows which mode it's in. They show LIBERO numbers on par with or above the 10-step teacher, ~3.3x lower latency, and a short real-robot demo.

Critical strengths

The problem is real and the fix is cheap. Most of the inference time in these models goes to the repeated action-expert calls, and this cuts them down to one with a pretty light fine-tune and basically no change at deployment. If it holds up, people would actually use it.

Critical weaknesses

Not clear what it buys over just running the model with one step. Naive one-step inference has the same speedup for free, and on π0.5 it's already within a couple of points of SnapFlow on a benchmark that's close to saturated. The π0 teacher is way below its published number, so beating it doesn't say much. I'd also expect at least one other few-step baseline (e.g. shortcut/MeanFlow-style fine-tuning, or just the teacher at 2-3 steps). Also, the 3.3x model-latency gain turns into only ~1.5x shorter episodes even under the synthetic timer in the LIBERO side-by-side videos (e.g. 14.4s vs 9.6s), and in one clip (t2200) SnapFlow actually takes longer.

The offline numbers are a bit misleading. Action MSE against ground truth naturally favors one-step / mean-like predictions (the teacher itself gets worse with more steps), so the big error reductions and the ablations built on them don't convince me much about actual policy quality.

Questions for authors
  1. How does the teacher do in closed loop at 2-3 steps, and how does a model that only gets the extra flow-matching fine-tuning (α = 1) do? If either is close to SnapFlow, that changes my view a lot.
  2. Any multi-seed or significance numbers for the LIBERO comparisons, especially SnapFlow vs. naive one-step?
  3. Do the one-step actions still vary across noise samples, or is the model just collapsing to a mean action?
3 comments
AuthorsOpenReview ↗

Response to Reviewer KUSc (1/2)

Your Q1 and Q2 identify the central issue, and the answer changes the paper's main claim. We report the paired results below, including those that do not favor SnapFlow; the general response gives the full tables.

W1 / Q2. What does SnapFlow buy over naive one-step inference, and is the comparison significant?

On LIBERO, nothing measurable. Under the paired protocol (same initial states and evaluation seed for every method, exact McNemar tests) the π0.5 teacher and its naive one-step sample both reach 98.25%, and the submitted-recipe student 96.25% and 97.50% for two training seeds (p = 0.10 and 0.63 against naive one-step). On the public π0 checkpoint with 3 evaluation seeds (n = 1200) the teacher reaches 1134, naive one-step 1127 and the revised student 1125 (p = 0.41 and 0.92). The submitted π0 teacher at 70.5% was an outlier that its own one-step inference beat by 7.25 points, so the +10.5 pp over it said little, as you noted; the revision replaces it by the public checkpoint and withdraws every "surpasses the teacher" statement.

Q1. Teacher at 2 to 5 steps, and the α = 1 control.

Teacher at 1, 2, 3, 4, 5 and 10 NFE on the public π0 checkpoint, paired, evaluation seed 1000, 100 episodes per suite (the 2-NFE run lacks LIBERO-Spatial, so its total is over three suites):

NFESpatialObjectGoalLongTotal
196999686377/400
2—989585278/300
3951009089374/400
4971009989385/400
5961009588379/400
1097999483373/400

The range across step counts, 373 to 385 of 400, is within single-seed noise (the 95% CI of a paired difference at n = 400 is about ±4 pp), and no step count beats one step. Over three evaluation seeds the 10-step and 1-step teachers differ by −0.6 pp (p = 0.55, n = 1200). The α = 1 control with exactly the same budget (30k updates, batch 4, frozen VLM, fp32 master weights), used as an ordinary flow model at 1 NFE, reaches 1101/1200 (91.8%), 2.8 pp below the untouched teacher (p = 0.005) and 2.2 pp below its naive one-step sample (p = 0.026). The submitted SnapFlow recipe with the two implementation fixes described in the general response reaches 1084/1200 (90.3%); the 1.4 pp between the two fine-tuned models is not significant (p = 0.18). So the control is "close to SnapFlow", as you anticipated, but both sit below the untouched teacher: on a converged LIBERO teacher any further fine-tuning at batch 4 costs points. The revised recipe described in the general response (frozen-teacher target, all gradient on the one-step branch, effective batch 32; not the method as submitted) recovers to 1125/1200 (93.75%), indistinguishable from the teacher (p = 0.41) and from naive one-step (p = 0.92). The revision states that on LIBERO a one-step policy gains nothing over naive one-step inference, because the teacher itself gains nothing from its extra steps (Section 4.2; recipe ablation in Appendix O; teacher step-count sweep in the same appendix).

Q2. Multi-seed and significance.

All π0 numbers above use 3 evaluation seeds × 400 paired episodes and exact McNemar tests; the tables in the general response give b/c-based p-values for every comparison, and Appendix D.2 of the revision lists the success counts per evaluation seed and suite for every π0 method. The π0.5 results (one evaluation seed, n = 400) are given above; none of the differences is significant. On LIBERO-Plus (400 paired perturbed variants) teacher, naive one-step and the revised student all score 263/400 (p = 1.0 for every pair; Appendix D.3). On PushT (see the general response; n = 200 paired episodes, one training and one evaluation seed; Appendix D.4) the differences to the teacher's own one-step sample are significant: teacher 10 steps vs 1 step −17.5 pp (b = 49, c = 14, p < 0.001); the submitted recipe with frozen-teacher target (batch 8) vs naive one-step +13.0 pp (b = 15, c = 41, p = 0.001), vs the teacher −4.5 pp (p = 0.35); the revised recipe vs naive one-step +11.5 pp (b = 15, c = 38, p = 0.002), vs the teacher −6.0 pp (b = 47, c = 35, p = 0.22). The differences among the two students and the FM-only control are not: +4.5 pp (b = 26, c = 35, p = 0.31) and +3.0 pp (b = 31, c = 37, p = 0.55) over the control, and −1.5 pp between the students (p = 0.81). The revised recipe also used half the samples and an eighth of the optimizer steps of the control (120k samples, 3,750 updates vs 240k, 30k), so its PushT number is not an equal-budget comparison, and we do not read it as an advantage.

AuthorsOpenReview ↗

Response to Reviewer KUSc (2/2)

Q3. Do one-step actions still vary with the noise, or is the model collapsing to a mean?

We measured it directly. On 80 LIBERO-Long training frames (8 per task) we drew the same 16 noise samples for every method and computed the per-element standard deviation of the predicted chunk across samples, in the normalized action space (the across-frame standard deviation of the ground-truth chunks is 1.03). The 10-step teacher's dispersion is 0.056, i.e. 5.4% of the data scale; no frame exceeds 0.13 and the per-task range is 0.044 to 0.068. The conditional action distribution π0 has learned on LIBERO is therefore close to a point mass, which is the mechanism behind the zero headroom above: a deterministic one-step prediction, which approximates the conditional mean, loses almost nothing. The naive one-step teacher retains 39% of the 10-step dispersion (0.022 vs 0.056, ratio of frame-averaged dispersions), the α = 1 control 40%, and the submitted-recipe SnapFlow student 50% (0.028; mean sample-to-sample distance 0.36 vs 0.27 for naive one-step), while the mean predictions of all variants are equally accurate (MSE to ground truth 0.0125 to 0.0158). So the student is partly mode-averaging on LIBERO, as is the teacher's own one-step sample, and there is little to preserve. On the PushT teacher the same diagnostic gives a 10-step dispersion of 10.1% of the data scale, about twice LIBERO's 5.4%, and its naive one-step sample retains only 32% of it; this is the regime in which the closed-loop headroom is 17.5 pp. The dispersion of the PushT students was not measured within the discussion period. This diagnostic is Appendix D.5 of the revision, and we drop offline MSE as evidence of policy quality.

W2. 3.3× model latency vs. about 1.5× shorter episodes; the t2200 clip.

You are right that the model-latency speedup does not carry over to episode time. The side-by-side clips use a synthetic timer: elapsed time = policy steps × the measured per-chunk latency, as stated in the supplementary README. This makes episode duration depend on how many chunks each policy needs; when SnapFlow needs more chunks, as in t2200, its clip runs longer. The clips were also selected from episodes where both policies succeed, so they do not reflect the full distribution. We did not complete a per-episode wall-clock measurement within the discussion period, so the revision makes no claim about episode time or throughput: it describes 3.3× as batch-one model-inference latency only, adds that naive one-step decoding gives the same speedup, so the case for distillation cannot rest on speed (Section 5), and states that no real-robot latency was measured (Section 4.5).

W3 / Q3. Offline MSE favors mean-like predictions.

We agree that MSE against a single ground-truth action rewards mean-like predictions; the fact that the teacher's MSE rises with more steps (Table 9 of the submission) is consistent with this, and much of the offline gap is shared by naive one-step decoding. The revision no longer uses offline MSE as evidence of policy quality (Section 4.3). While re-checking these numbers we also found an error in the paper: the 500 "held-out" samples in Tables 2, 3 and 9 of the submission are the first frames of the LIBERO demonstration set, which both the teacher and SnapFlow were trained on. They are not held out, and the revision says so everywhere they appear. Across runs with different noise draws, the 10-step teacher's mean MSE on this set varies by about ±20%; the differences in the submitted Table 3 lie within that range, so the revision presents that ablation as an exploratory diagnostic only (Appendix P) and bases the ablation of the training recipe on paired closed-loop results instead (Appendix O).

Summary of changes for this review: paired multi-seed results with McNemar tests replace the unpaired π0.5 numbers and the outlier π0 teacher; teacher at 2 to 5 NFE and the α = 1 control added, on LIBERO and PushT; a headroom table for every base policy and benchmark; the dispersion diagnostic; offline MSE demoted to a diagnostic and the "held-out" wording corrected; the episode-time claim removed; the two implementation errors disclosed and fixed; the revised recipe adopted with the submitted recipe reported as an ablation.

Thanks for the additional experiments and corrections. These address my questions, but my main concern remains. On LIBERO and LIBERO-Plus, the revised SnapFlow offers no clear improvement over naïve one-step inference, which gives the same model-inference speedup. On PushT, it improves over naïve one-step inference, but the results do not yet show a clear advantage over continued flow-matching training.

I appreciate the authors’ careful follow-up and the narrower claims. However, I am still not convinced that the results demonstrate enough benefit from distillation to support acceptance.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General response (1/2)

All closed-loop numbers below are paired: the same task, initial state and evaluation seed for every method, with exact McNemar tests on the paired episodes. On the public π0 LIBERO checkpoint we use 3 evaluation seeds × 4 suites × 100 episodes (n = 1200) per method. We report every result we obtained, including those that do not favor SnapFlow. Revised text in the uploaded PDF is marked in blue.

Headroom. For a given base policy we call headroom the paired difference between its 10-step success and its own naive one-step success (one Euler step from pure noise, which approximates the conditional mean action). It measures how much a one-step student can recover from the teacher's multi-step sampling. On LIBERO this headroom is zero for both base models: π0, 94.5 vs 93.9 (−0.6 pp, 95% CI [−2.2, +1.0], p = 0.55, n = 1200); π0.5, 98.25 vs 98.25 (n = 400). There is nothing to recover, so a one-step method cannot be shown to add anything on this benchmark, and we withdraw the "exceeds the teacher" wording. The π0 teacher in the submitted Table 1 was an outlier whose 10-step inference was worse than its own one-step inference (70.5 vs 77.75, p = 0.006); the +10.5 pp we reported over it is that gap plus a non-significant +3.25 pp over naive one-step. We replace that row by the public checkpoint. The revision adds a headroom table for every base policy and benchmark we use (Appendix D.1).

Two implementation issues affected all submitted training runs. First, the training schedule clamped t ≤ 0.98 while one-step inference queries t = 1; with the clamp the one-step student scores 168/300 on LIBERO-Long, without it 242/300 (teacher 259/300). Second, under dtype=bfloat16 the action expert was stored and updated in bf16, and at our learning rate about 79% of its weights never changed over 30k updates. Both are fixed in the revision (no upper clamp; fp32 master weights with bf16 compute; Appendix O).

Corrected LIBERO results (π0, 1 NFE unless stated, n = 1200, paired; Δ in pp):

MethodSuccessΔ vs teacher 10 NFE (p)Δ vs naive 1 NFE (p)
Teacher, 10 NFE1134 (94.5)—+0.6 (0.55)
Naive one-step, no training1127 (93.9)−0.6 (0.55)—
Submitted recipe, bf16, clamp fixed1096 (91.3)−3.2 (0.001)−2.6 (0.003)
Submitted recipe, both fixes (online target, batch 4)1084 (90.3)−4.2 (<0.001)−3.6 (<0.001)
FM-only fine-tune, same budget (α = 1)1101 (91.8)−2.8 (0.005)−2.2 (0.026)
Submitted recipe with frozen-teacher target, batch 41079 (89.9)−4.6 (<0.001)−4.0 (<0.001)
Submitted recipe, both fixes, effective batch 321116 (93.0)−1.5 (0.11)−0.9 (0.36)
Revised recipe (defined below)1125 (93.75)−0.75 (0.41)−0.2 (0.92)

Three facts follow. Any further fine-tuning of this converged teacher at batch 4 costs 2 to 4 pp at one step, the FM-only control included, so most of the submitted recipe's loss is a fine-tuning cost, not a distillation cost (recipe vs control: 1.4 pp, p = 0.18). Effective batch 32 by gradient accumulation, with the same 120k training examples, recovers most of it. The revised recipe, which keeps the target network frozen at the pretrained teacher and trains only the branch used at deployment, is the only fine-tuned student statistically indistinguishable from the 10-step teacher. It does not exceed the teacher or the naive one-step sample; on LIBERO there was nothing to recover from the teacher's sampling.

The last row is not the method as submitted. The revised recipe changes four things relative to the submitted one: the target network is the frozen pretrained teacher rather than the online self-target; every distillation row uses destination s = 0 with weight 1, instead of half the rows at s = t with weight 0.1; half of those rows are placed at t = 1, the query made at inference; and the effective batch is 32 instead of 4. The submitted recipe with only the two bugs fixed is the 1084/1200 row. The revised paper adopts the revised recipe as SnapFlow and reports the submitted recipe, with the fixes, as an ablation (Table 1 and Appendix O of the revision; the counts per evaluation seed and suite are in Appendix D.2).

AuthorsOpenReview ↗

General response (2/2)

Out-of-distribution headroom. On LIBERO-Plus we evaluated a fixed paired set of 400 perturbed variants (80 per dimension: background, language, lighting, layout, sensor noise; one episode per variant). The 10-step teacher and its naive one-step sample both succeed on 263/400 (65.8%; paired difference 0.0 pp, p = 1.0). The revised one-step student also reaches 263/400 (p = 1.0 against both), and no dimension differs significantly. Sensor noise is the only dimension ordered in the expected direction (teacher 9/80, student 6/80, naive 3/80; p ≥ 0.38). The per-dimension table is Appendix D.3 of the revision.

Where multi-step sampling does matter. We added PushT (lerobot/pusht; the LIBERO π0 weights fine-tuned with flow matching for 30k steps at batch 8 serve as teacher; closed-loop gym-pusht, 200 paired episodes, one training seed and one evaluation seed; success = coverage ≥ 95%; Appendix D.4). There the demonstrations are multimodal and the same teacher loses 17.5 pp when sampled in one step instead of ten (30.0% vs 12.5%, b = 49, c = 14, p < 0.001) and 0.145 of final coverage (0.811 vs 0.666); at two steps it reaches 22.5%. Starting from this teacher, three one-step models recover most of that gap in success: the submitted recipe with frozen-teacher target (batch 8) 25.5% (+13.0 pp over naive one-step, p = 0.001; −4.5 pp vs the teacher, p = 0.35); the revised recipe 24.0% (+11.5 pp, p = 0.002; −6.0 pp vs the teacher, p = 0.22); and a control that only continues flow-matching training for 30k steps at batch 8 and is then sampled in one step, 21.0% (+8.5 pp, p = 0.036; −9.0 pp vs the teacher, p = 0.06). The three are statistically indistinguishable from each other: frozen-target student vs control +4.5 pp (p = 0.31), revised recipe vs control +3.0 pp (p = 0.55), revised vs frozen-target −1.5 pp (p = 0.81). None of them recovers coverage: all three stay significantly below the teacher (revised recipe −0.092, paired t = −3.7; frozen-target −0.107; control −0.070), and the control recovers coverage at least as well as the students. The budgets are also unequal: the control and the frozen-target student each saw 240k samples over 30k optimizer steps, whereas the revised recipe saw 120k samples over 3,750 optimizer steps (micro-batch 4, accumulation 8, as on LIBERO); an equal-budget run of the revised recipe is not included, so we do not read its number as an advantage. The absolute rates are those of a 7-DoF LIBERO policy fine-tuned briefly on a 2-D task and are not comparable to models built for PushT.

Revised claims. SnapFlow converts a pretrained flow VLA into a one-step policy with a frozen VLM, no change at deployment and the lower model latency reported in Table 8 of the submission (Appendix H of the revision). On LIBERO and LIBERO-Plus the one-step student matches the 10-step teacher within noise, as does the teacher's own naive one-step sample. Where the headroom is positive, as on PushT, the one-step students recover most of the teacher's 10-step success, but so does continued flow-matching training, and the three are statistically indistinguishable, so we claim no contribution of the distillation target beyond plain fine-tuning. We add a headroom table for every base policy and benchmark we use (Appendix D.1), so that the attainable gain is visible before the student's number. We no longer use offline MSE as evidence of policy quality, present the real-robot material as a qualitative check only, and describe the 3.3× figure as batch-one model latency only, without any claim about episode time, which we did not measure under a controlled protocol. A Shortcut Models baseline on the same backbone was not completed within the discussion period and is not reported.

Other corrections. The 500 "held-out" offline samples of the submission are frames from the training demonstrations; the revision says so and uses them only as a diagnostic (Section 4.1, Appendix F). The α = 1 row of the submitted offline ablation queried the destination-time input at a value it was never trained on and is marked invalid (Appendix P); the closed-loop FM-only control above replaces it. The n_act = 10 point of the action-horizon sweep now uses the sweep's own recorded values (teacher 89, SnapFlow 91) and the sweep is labelled single-seed (Appendix J). Figure 2(a) is fixed so that all four series are visible. The four LIBERO suites are named in Table 1, the distillation data and budget are stated in Section 4.1, training time is given for fp32 (about 12 h) and bf16 (about 2.6 h), Theorems 1–3 are moved to Appendix A as background from prior work together with the relation between the corrected objective (Eq. 8) and the two-step target, and Related Work discusses the one-step VLA methods raised by Reviewer Lad6.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.