acceptodds
Under review as a conference paper at ICLR 2027

Flow Matching Is Not a Contractive Class for Vision-Language-Action Policies

Abstract

Vision-language-action (VLA) policies are increasingly employing flow matching for robot action generation. The process starts from random noise and produces an action over a few steps. It is currently believed that flow matching intrinsically contracts noise perturbations and this leads to robust VLA policies. Recent approaches often build upon this idea. We show that this belief is actually not correct. Contraction and expansion of flow in VLAs are mainly tied to the model architecture, not to the flow matching process. We establish this with a rigorous treatment of four diverse state-of-the-art VLAs. We achieve a clear split of two contractions and two expansions, which is not caused by flow matching, the direction of the flow, or the number of steps. Two policies sharing similar process design choices still land on the opposite sides, isolating the action-generation module as the primary cause. Our analysis further reveals two crucial insights. (a) Contraction in a deployed policy must be judged over the entire flow, not one step at a time, which is the current norm and can be misleading. We provide results supporting expansion at every step for a policy whose flow as a whole still contracts. (b) For flow-matching VLAs, robustness must not be inferred from the flow and needs to be measured independently because model robustness may not align with contraction. Our supporting evidence identifies a flow expanding policy as the most robust one. Our source code will be made public after acceptance.

Primary area:
applications to robotics, autonomy, planning
Keywords:
vision-language-action models, flow matching, robot policies, contraction analysis, robustness, generative models

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_zMjUOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper examines whether flow-matching VLA policies contract perturbations in their initial sampling noise. It measures the largest singular value of a composed sampler Jacobian across four policies and reports both contractive and expansive behavior. It also distinguishes sensitivity to sampling noise from sensitivity to observations and compares end-to-end contraction regularization with a per-step alternative.

Critical strengths

The distinction between noise-to-action sensitivity and observation-to-action sensitivity is useful. It helps clarify what a contraction measurement can establish about a deployed policy. The four-policy study includes finite-difference checks, sampler analyses, and supplementary measurements. These empirical observations could be valuable independently of the stronger theoretical claims.

Critical weaknesses

The interpretation of Proposition 3 and Appendix A conflates continuous-time contraction with contraction of a discrete Euler step. In the stated counterexample, both Euler factors already have operator norm approximately 1.526, while their product has norm approximately 2.329. Thus, the example demonstrates that discretization can break contraction; it does not demonstrate that genuinely contractive steps compose into an expansive map. Appendix F also applies a raw symmetric-Jacobian penalty to policies with negative sampling time steps, without clearly accounting for the integration direction. These issues undermine the central comparison against per-step contraction analysis.

The claimed exactness and validation of the measurement need reconciliation. Section 3.1 describes an exact deployed-sampler Jacobian, whereas Appendix D states that the reported object is a product of active-block restrictions. These generally differ when perturbations can propagate through padded dimensions. The supplementary full-Jacobian checks are helpful but do not establish equivalence throughout the evaluation. In addition, Section 4 says all models pass the validation protocol, while Table 3 marks nonlinear-predictiveness evidence as unavailable under the common protocol for two models.

Questions for authors

Do you agree that the Appendix A example establishes a discretization failure rather than failure of continuous-time contraction? Which central conclusions remain under this interpretation? How is the Arm C penalty derived for policies using negative time steps? Is the velocity Jacobian reoriented before applying the penalty? Are padded dimensions clamped at every step of the deployed sampler? If not, how does the product of active-block Jacobians account for paths through those dimensions?

4 comments
AuthorsOpenReview ↗

Response to Reviewer zMjU (1 of 3)

Response to Reviewer zMjU (1 of 3).

We thank the reviewer for a precise and careful review. The reviewer notes that the empirical measurements stand independently of the theoretical framing, and the new measurements below bear this out. We answer in three comments. This one answers Q1 and Q2, on the Appendix A example and on which conclusions remain. The second answers Q3 and Q4, on the Arm C penalty. The third answers W2, Q5 and Q6 on exactness, padded dimensions and Table 3, and lists all changes to the paper.

Q1, does the Appendix A example establish a discretisation failure? We agree that it does. Each Euler factor I + J_k in the example has operator norm 1.526. The example therefore shows that a step with λ_max(sym J_k) < 0 need not be a contraction once discretised at dt = 1. It says nothing about composing contracting maps, since two maps of norm below 1 cannot compose to a norm above 1. Proposition 3 and Appendix A will be restated as a discretisation result (C4).

The discretisation gap is the one that matters for a deployed sampler, and we have now measured it. The deployed policies are Euler samplers with dt = −0.1 or +0.25, while the field's per-step criterion λ_max(sym J_k) ≤ 0 is a continuous-time statement. Whether the continuous criterion controls the discrete step is therefore the question to check rather than assume. With the orientation made explicit (below), the per-step bound B = exp(Σ_k |dt| μ_k) upper-bounds the measured end-to-end gain on 960/960 draws of the paper's panel. On these samplers the Euler term never breaks the certificate.

Q2, which central conclusions remain? All four central conclusions remain. Three are not affected by the review, and the fourth is sharpened by it.

  1. Flow matching is not a contractive class (Sec. 5.1, Fig. 2a). Unchanged, and now verified on the deployed samplers themselves on 960/960 draws (comment 3).
  2. Generation-flow geometry does not predict observation robustness (Sec. 5.3, Table 2). Unchanged, and not affected by any point in the review.
  3. Contraction can be engineered but not for free (Sec. 5.4, App. F). Unchanged with the re-oriented penalty (comment 2).
  4. The per-step criterion versus the end-to-end gain (Sec. 5.2, Table 1, abstract insight (a), contribution 2). Sharpened. Read in the sampler's own integration direction, the per-step criterion supports the end-to-end measurement rather than competing with it; the derivation and the result follow.

The orientation. For a sampler that integrates with dt < 0 the step is I − |dt| J_k, so the controlling end of the spectrum is λ_min(sym J_k), not λ_max. We re-derived the per-step bound with the sign of dt explicit, so that μ_k = λ_max(sym J_k) for dt > 0 and μ_k = −λ_min(sym J_k) for dt < 0. We then re-evaluated it from the saved per-knot spectra of all 960 draws.

ModeldtDraws certified B < 1Draws with σ_max < 1Falsely certified (B < 1, σ_max ≥ 1)
GR00T+0.25160/160160/1600
π0.5−0.1292/320305/3200
SmolVLA−0.10/1600/1600
π0−0.10/3200/3200

The result. With orientation, the per-step criterion certifies π0.5's contraction on 292 of its 305 contracting draws and never certifies an expanding draw. The disagreement reported in Table 1 and insight (a) came from reading λ_max(sym J_k) on dt < 0 samplers. Once oriented, the one-sided per-step certificate is sufficient and conservative on the deployed samplers (13 contracting π0.5 draws are not certified, none is falsely certified). The end-to-end gain remains the quantity that decides the verdict and the one to report. This is the methodological point the revised paper makes.

In the revision (C4), Table 1, Sec. 5.2, insight (a) and contribution 2 are rewritten in these terms. The gate (ii) self-test is re-described, since its one-sided disagreement is also the orientation effect. The Related Work sentence on a policy where the two verdicts disagree is removed.

AuthorsOpenReview ↗

Response to Reviewer zMjU (2 of 3)

Response to Reviewer zMjU (2 of 3).

This comment answers Q3 and Q4 on the Arm C penalty of Appendix F.

Q3 and Q4, how was the Arm C penalty derived, and was the Jacobian reoriented? In the submission it was not. Arm C penalised Σ_k relu(λ_max(sym J_k)) on the velocity Jacobian at the sampler's knots. This is the CDP-class object as written for forward integration, while SmolVLA and π0 integrate with dt = −1/K. For a backward step I − |dt| J_k the per-step log-norm is λ_max(sym(−J_k)) = −λ_min(sym J_k). The submitted penalty therefore constrained the wrong end of the spectrum, as the reviewer inferred. We re-derived it with the orientation explicit as relu(λ_max(sym(sign(dt) J_k))). We then re-ran Arm C on the identical recipe (same training entries, steps, seed, penalty observations and held-out flows).

Re-oriented, the one-sided penalty contracts as well, and the paper's conclusion about the cost of contraction is unchanged.

Arm (σ_max median on 10 held-out flows)SmolVLA, steps 50 / 100 / 150π0, steps 50 / 100
Baseline1.8673.041
A, FM only1.405 / 6.058 / 1.4281.251 / 3.029
B, two-sided end-to-end0.713 / 0.850 / 0.9731.013 / 0.992
C, one-sided as submitted0.973 / 0.906 / 1.0561.611 / 2.752
C, one-sided re-oriented0.931 / 0.763 / 0.7501.010 / 1.018

On SmolVLA the re-oriented Arm C contracts at every checkpoint on 10/10 held-out flows, and it ends deeper than Arm B. On π0 it reaches the boundary (2/10 flows below 1), level with Arm B at 0.99 to 1.01. Training drove its oriented per-step log-norm to the penalty threshold of zero and stopped there. A per-step bound at zero leaves the O(dt²) Euler term of the step uncontrolled, which is the same discretisation point as Q1.

The sentence "only the two-sided object contracts; the one-sided object fails on both architectures" (Sec. 5.4, App. F) is therefore replaced. Appendix F now compares an end-to-end penalty with a correctly oriented per-step penalty, and both induce contraction. The conclusion that carries Sec. 5.4 is untouched and now rests on both arms. Every arm that contracts fails the pre-registered non-degeneracy gates, B and re-oriented C alike. Observation coupling falls to 0.09 to 0.17 of baseline. Contraction can be engineered, but not without this cost, and the tradeoff null stands (C5).

AuthorsOpenReview ↗

Response to Reviewer zMjU (3 of 3)

Response to Reviewer zMjU (3 of 3).

This comment answers W2, Q5 and Q6 on the exactness of the witness, the padded dimensions and Table 3. It then lists all changes to the paper.

Q5, are padded dimensions clamped at every step? They are not. Every deployed sampler evolves all 32 latent dimensions at each step and slices the 7 active DoF only at the output. The witness varies the active block with padding held fixed. It is therefore the active-block restriction of the full propagator, as Sec. 3.3 and App. D state. Proposition 1 is exact for the full 32-dimensional state; the reported witness is its restriction, and Sec. 3.1 should not call it exact.

Q6, how does the restriction account for paths through padding? It does not account for them by construction. We therefore measured them directly on the paper's panel, with the same seeds and draws. For the contractors we computed the full 32-dimensional deployed Jacobian on every draw. For the expanders we computed the exact Jacobian-vector product of the full sampler along the witness's top direction, which lower-bounds the deployed gain and suffices to decide the sign.

The witness is the restriction of an exact Jacobian, and on every draw of the panel it reads the same sign as the full deployed map.

Table. The full deployed sampler against the active-block witness, on every draw of the paper's panel.

ModelVerdictDraws on the witness's side of 1Full deployed gain
SmolVLAexpands160/160at least 1.33 on every draw
π0expands320/320at least 2.21 on every draw
π0.5contracts320/320within 1.1% of the witness
GR00Tcontracts160/1600.76 to 1.01 of the witness, median 0.92

The active columns carry 96.7% to 99.9% of the deployed gain, so paths through padding carry at most 3.3% and never change a verdict. The witness is within 1.1% of the deployed gain for π0.5 and conservative for GR00T. Sec. 3.1 will describe the witness as the active-block restriction of the exact Jacobian, within 1.1% of the deployed gain for π0.5 and conservative for GR00T. The scope paragraph of Sec. 3.3 moves next to it (C1).

Table 3. The two entries marked † are now witnessed on the common footing. The same audit records the full nonlinear deployed gain ρ along the witness direction at ε ∈ {3e-2, 1e-2, 3e-3, 1e-3, 3e-4} on every draw. ρ lands on the witness's side of 1 at all five ε for all 960 draws. At ε = 1e-3 the median deviation of ρ from the witness is 0.08% (SmolVLA), 0.36% (π0), 0.02% (π0.5) and 9.8% (GR00T, the conservative gap above). Its deviation from the exact Jacobian-vector product is 0.01 to 0.02% for all four. Sec. 4 will state gate (iii) at ε ≤ 0.03, the range tested on this footing. Table 3 is completed (C2).

Changes to the paper (revised PDF, within the 10-page limit). C1 Sec. 3.1 exactness wording, scope statements consolidated, deployed-sampler audit in App. D. C2 Table 3 completed, gate (iii) at ε ≤ 0.03. C4 and C5 as stated in comments 1 and 2. C7 presentation, which the reviewer rated lowest. The sign of dt and the orientation of every per-step quantity are stated once in Sec. 3.2 and used consistently. Each quantity is introduced in words before its symbol. Derivations not needed for the main argument move to the appendix. Changes for the other reviewers (premise sources, multimodality probe) are in the general response.

Both concerns have been addressed by measurement rather than by argument, and the central result is stronger than it was at submission. If the reviewer agrees, we would be grateful if this were reflected in the soundness assessment, and we welcome any further question.

AuthorsOpenReview ↗

Response to Reviewer zMjU (summary)

As the discussion period closes on 18 November, we summarise briefly how the two main concerns raised in the review were addressed by measurement, and where the corresponding changes appear in the revised PDF.

W1. Read in each sampler's own integration direction, the per-step criterion certifies contraction for GR00T (160/160) and π0.5 (292/320), certifies none for the expanders, and is never falsely satisfied. Proposition 3, App. A, Table 1 and Sec. 5.2 are restated in these terms (C4). The re-oriented Arm C penalty also induces contraction, and the conclusion on the cost of contraction is unchanged (App. F, C5).

W2. The witness is the active-block restriction of the exact Jacobian. On all 960 draws it reads the same sign as the full deployed map, with paths through padding carrying at most 3.3% of the gain. Sec. 3.1 now describes it in these terms and the audit is in App. D (C1). Table 3 is completed on the common footing (C2).

We hope this summary is helpful and would be glad to answer any remaining question before the discussion closes.

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

This paper investigates whether flow-matching-based Vision-Language-Action (VLA) policies inherently exhibit contraction during action generation. The authors examine the contraction properties of deployed VLA samplers by computing the largest singular value of the noise-to-action Jacobian. They further argue that the conventional one-sided per-step criterion is neither necessary nor sufficient to characterize contraction of the complete discrete sampling process. Experiments on four VLA policies using LIBERO show that SmolVLA and π0 exhibit expansion, while GR00T-N1.5 and π0.5 exhibit contraction. Notably, π0 and π0.5 share a similar backbone, sampling direction, and step count despite exhibiting opposite behaviors. The authors also show that generation-flow contraction does not necessarily correspond to robustness against observation perturbations. Additional experiments demonstrate that a contraction penalty applied to the composed sampler can enforce contraction, but at the cost of reduced observation responsiveness. Overall, the paper argues that contraction is not an intrinsic property of flow matching but depends on the learned policy and its action-generation design.

Critical strengths
  1. The paper evaluates four representative VLAs on LIBERO, revealing clear differences in contraction and expansion, particularly between π0 and π0.5.

  2. The Jacobian-based measurements are supported by theoretical derivations, finite-difference validation, and additional numerical checks, providing a solid basis for analyzing deployed VLA samplers.

Critical weaknesses
  1. The paper's central motivation relies on the claim that flow matching is widely believed to be inherently contractive. However, the cited literature primarily studies contraction under specific assumptions or explicitly enforces it through additional regularization. It remains unclear whether the claimed misconception is actually widespread in the VLA community.

  2. The paper primarily applies established mathematical tools to existing VLA policies. More importantly, the observed expansion may reflect meaningful multimodal action generation rather than undesirable sensitivity, while contraction may limit action diversity. Although Appendix F discusses this trade-off through a 1D example, the paper does not investigate whether these phenomena explain the results observed in actual VLAs.

Questions for authors
  1. Could the authors clarify which prior works support the claim that flow matching is widely assumed to be intrinsically contractive, and distinguish this assumption from methods that explicitly enforce contraction?

  2. How should the observed contraction/expansion differences be interpreted in terms of VLA behavior? In particular, could expansion reflect transitions between valid action modes rather than undesirable sensitivity? Is there any evidence from the evaluated VLAs supporting this interpretation?

5 comments
AuthorsOpenReview ↗

Response to Reviewer woif (1 of 2)

Response to Reviewer woif (1 of 2).

We thank the reviewer for the careful review. The reviewer recognises the clear split between π0 and π0.5 and the validation behind the measurement, both central to the paper. We answer in two comments. This one answers W1/Q1, the premise and the distinction between assuming contraction and enforcing it. The second answers W2/Q2 with a new experiment on the evaluated VLAs and lists the changes to the paper.

W1/Q1. The reviewer's distinction is the right one, and the paper rests on it. Sec. 2 frames the premise as inherited from control and sampler theory and enforced by CDP, and as untested on deployed VLAs. It reads the enforcement literature as enforcement. The revision adds the sources in which the assumption itself is stated, so that both sides of the distinction are cited and named.

The assumption is stated in the applied VLA literature, about the very policies we test. Nahim et al. [1] appeared after the submission deadline. This survey lists as an advantage of the deterministic flow that "slight perturbations in the latent noise do not result in wildly different action sequences." This is a statement about the noise-to-action gain, which is the quantity σ_max(Φ) we measure. ACG [2] describes "generative flow-based VLA models (e.g., GR00T-N1, π0, SmolVLA), which offer improved stability and expressiveness." DiffusionVLA [3] credits "the robustness of high-frequency action generation offered by diffusion models." RobustVLA [4] attributes π0's "stronger intrinsic robustness" over its autoregressive variant to "the diffusion-based action head." Psiris et al. [5] write that π0.5 "casts action generation as flow matching for stable continuous control." None of these tests the property.

Where contraction is needed it is enforced, which confirms it does not come for free. CDP [6] penalises the per-step quantity λ_max(sym J_k), which is the one-sided criterion our paper evaluates against the end-to-end gain (Sec. 5.2). It does so "to induce contractive behavior in the diffusion sampling dynamics," because "contraction pulls nearby flows closer to enhance robustness." Tang and Zhao [7] require that "the backward process of the DPM shall be contractive" and note that in standard models errors "may be amplified." Ding et al. [8] regularise their flow-matching policy because plain flow matching does "not guarantee that the flow converges stably." These works share the premise that contraction is the route to robustness, and each builds it in.

No theorem decides it for action data, so it had to be measured, and we are the first to do so on deployed policies. Contraction of the probability-flow ODE is proven only under strong log-concavity [9]. Without it the dynamics are "not contractive, and hence small errors quickly accumulate and are magnified" [10]. Multimodal action distributions are not log-concave, so no result decides the sign for a VLA in either direction.

The paper addresses the gap between the first group and the other two. Deployed VLAs are credited with a property that the methods needing it have to engineer and that theory does not grant, and no one had measured whether they have it. The abstract and Sec. 1 will state the premise in these terms and cite the first group, where the belief is stated (C3). Independent evidence that the belief circulates comes from a 2026 preprint on attacking flow-matching VLAs [11]. It opens by noting that they "have been reported to resist adversarial perturbations that readily fool autoregressive VLAs" and finds this "largely illusory."

[1] N. Z. Nahim, F. N. Khan, P. Ju. A Survey of Flow Matching in Reinforcement Learning. arXiv:2610.16275, 2026.

[2] M. Park et al. ACG: Action Coherence Guidance for Flow-based Vision-Language-Action Models. ICRA, 2026.

[3] J. Wen et al. DiffusionVLA: Scaling Robot Foundation Models via Unified Diffusion and Autoregression. ICML, 2025.

[4] J. Guo et al. On Robustness of Vision-Language-Action Model against Multi-Modal Perturbations. ICLR, 2026.

[5] A. Psiris et al. Foundation Models in Robotics: A Comprehensive Review of Methods, Models, Datasets, Challenges and Future Research Directions. arXiv:2604.15395, 2026.

[6] A. Abyaneh et al. Contractive Diffusion Policies. ICLR, 2026.

[7] W. Tang, H. Zhao. Contractive Diffusion Probabilistic Models. SIAM Journal on Imaging Sciences, 2026.

[8] H. Ding, N. Jaquier, J. Peters, L. Rozo. Fast and Robust Visuomotor Riemannian Flow Matching Policy. IEEE Transactions on Robotics, 2025.

[9] X. Gao, L. Zhu. Convergence Analysis for General Probability Flow ODEs of Diffusion Models in Wasserstein Distances. AISTATS, 2025.

[10] S. Chen, S. Chewi, H. Lee, Y. Li, J. Lu, A. Salim. The Probability Flow ODE is Provably Fast. NeurIPS, 2023.

[11] H. Tae, J. S. Lee. DRIFT: Derailing Denoising Trajectories of Flow-Matching VLAs with Adversarial Patch Attack. arXiv:2608.03207, 2026.

AuthorsOpenReview ↗

Response to Reviewer woif (2 of 2)

Response to Reviewer woif (2 of 2).

This comment answers W2/Q2 with a new experiment on the evaluated VLAs. It then addresses the contribution and lists the changes to the paper.

Q2, is expansion a transition between valid action modes? We tested this directly, with design and verdict rules fixed before any data. At each of 40 states per model we draw 256 actions from the deployed sampler. A pre-registered detector tests each cloud for distinct modes (PCA, Gaussian mixture by BIC, Ashman's D ≥ 2, Hartigan dip test p < 0.01, calibrated on each model's own clouds). We then measure the gain at the states with a single mode.

Table. Is expansion mode switching? 40 states per model on libero_10 (20 at frame 0 as in the paper and 20 mid-episode). 256 sampled actions and 4 Jacobian draws per state. If it were, the expanders would show several modes and would stop expanding where there is only one. Neither expander shows more than one mode at any state, and both expand at every state.

ModelPaper verdictStates with more than one mode (of 40)σ_max at single-mode states (median)Draws with σ_max > 1 (of 160)
SmolVLAexpands02.65160 (smallest 1.18)
π0expands03.26160 (smallest 2.65)
GR00Tcontracts10.3310
π0.5contracts00.286

Every expander draw expands (95% CI of the median [2.17, 3.27] and [3.11, 3.66], Holm-corrected p = 0.001 against median ≤ 1). This holds under all six pre-registered variants, including a loose detector that catches closer modes, labels 14 SmolVLA states multimodal and still gives median 3.30 at the rest. The effect is not confined to small perturbations. Along the top direction the finite-displacement gain stays at 2.1 to 3.2 (SmolVLA) and 2.5 to 3.2 (π0) out to three noise standard deviations. Expansion in the expanders is within-mode sensitivity to the sampling noise, not a transition between modes.

The mechanism the reviewer describes is present in the contractors. GR00T's one multimodal state is a mid-episode gripper-timing decision. There, draws deep inside a mode have gain 0.57 and draws near the boundary have gain 8.87 (two draws each, illustrative). Moving along the top direction switches modes in 75% of line searches, against 25% for random directions. In a contracting policy, local expansion marks a decision between actions, and contraction does not remove the ability to switch. The reviewer's intuition that contraction may limit diversity matches our App. F finding that enforced contraction reduced observation coupling (the tradeoff null).

What this means for VLA behaviour. The gain compares how much the executed action depends on the sampling noise against how much it depends on the scene. The ratio of action spread over noise draws (fixed observation) to spread over observations (fixed draw) is 1.33 for π0, 0.49 for SmolVLA, 0.19 for GR00T and 0.09 for π0.5. In π0 the noise moves the action more than changing the scene does, within one mode. In the contractors the observation decides the action and the noise matters only at decision points. That is the behavioural meaning of the split, and why robustness must be measured rather than read off the flow (Sec. 5.3).

Two scope statements follow. GR00T's 80/80 contraction (Fig. 2a) is at frame 0, the paper's protocol; mid-episode it contracts on 17/20 states and π0.5 on 19/20. The App. F corollary that expanders are more multimodal is not supported at this detector, since neither group shows modes, and App. F will state the measured result (C6).

On the contribution (W2). The linear algebra is standard, and the paper claims no novelty for it. The contribution lies in the object it is applied to and in what the measurement shows. No prior work measures the end-to-end gain of a deployed VLA sampler, and doing so establishes facts that could not be obtained otherwise. The sign is set by the action head, not by flow matching (the π0 versus π0.5 pair the reviewer highlights). The per-step criterion the field certifies and trains against must be oriented to the integration direction and is conservative even then (general response). The sign does not predict observation robustness (Table 2). The probe above adds that expansion is within-mode. The central result is now also verified on the deployed samplers (960/960 draws; general response).

Changes to the paper. C3 as above. C6 the probe enters the appendix as a new section with the table above. The frame-0 scope is stated in Sec. 5.1 and the App. F corollary is replaced by the measured result. C7 each quantity is stated in words before its symbol in Sec. 1 and 3, and derivations not needed for the main argument move to the appendix.

If these answers resolve the two weaknesses, we would be grateful if the reviewer would reflect that in their assessment, and we welcome any further question.

AuthorsOpenReview ↗

Response to Reviewer woif (summary)

As the discussion period closes on 18 November, we summarise briefly how the two concerns raised in the review were addressed, and where the corresponding changes appear in the revised PDF.

W1/Q1. The belief that flow-matching action heads give stable, robust action generation is stated in peer-reviewed work on the policies we test, while the works that need contraction enforce it in training. The revised abstract and Sec. 1 now cite both groups and draw this distinction (C3).

W2/Q2. A pre-registered probe at 40 states per model finds no distinct action modes for the expanders, yet all 160 of their draws expand. Expansion is therefore within-mode sensitivity rather than mode switching. The probe and its table are in the new App. G, and the App. F corollary is revised accordingly (C6).

We hope this summary is helpful and would be glad to answer any remaining question before the discussion closes.

Reviewer_pE56OpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper investigates how flow-matching vision-language-action (VLA) policies contract or amplify perturbations in their initial action noise. Experiments on four models on LIBERO show that π0 and SmolVLA exhibit expansion, while π0.5 and GR00T-N1.5 predominantly exhibit contraction, indicating that flow matching does not inherently guarantee contraction. By comparing per-step criteria with end-to-end measurements, the paper emphasizes evaluating the complete generation process. Further experiments show that noise contraction does not directly predict sensitivity to visual perturbations, and that training to promote contraction may reduce responsiveness to observations.

Critical strengths
  1. Examines the fundamental question of whether flow matching inherently induces contraction in VLA policies.

  2. Provides extensive experiments across multiple architectures, supported by controls and validation checks.

  3. Clearly distinguishes sensitivity to initial action noise from robustness to visual perturbations.

  4. Offers a practical end-to-end method for measuring perturbation amplification during action generation.

Critical weaknesses

The paper does not sufficiently substantiate its central premise that prior research widely regards flow matching as inherently contractive and therefore a source of VLA robustness. Its key reference, CDP, introduces an additional regularizer to promote contraction in diffusion policies rather than claiming that flow matching is inherently contractive. This gap between the cited evidence and the position being challenged weakens the paper’s motivation.

Questions for authors
  1. Could the authors provide direct evidence from prior work to substantiate the claim that “it is currently believed that flow matching intrinsically contracts noise perturbations and this leads to robust VLA policies”?
4 comments
AuthorsOpenReview ↗

Response to Reviewer pE56 (1 of 2)

Response to Reviewer pE56 (1 of 2).

We thank the reviewer for the careful review. The reviewer recognises the cross-architecture study with its controls and validation, and the separation of noise sensitivity from observation robustness. We answer in two comments. This one gives the direct evidence requested in W1/Q1. The second states what it establishes and the changes to the paper.

W1/Q1. The belief is stated in peer-reviewed work, about the quantity we measure. The reviewer reads CDP correctly. So does our Sec. 2, which states the premise as inherited from control and sampler theory and enforced by CDP, and as untested on deployed VLAs. The abstract and Sec. 1 compress this to "intrinsically contracts," which is sharper than any single source. We will align them with Sec. 2 and cite the sources below.

  1. The belief is stated about the latent noise, the exact quantity σ_max(Φ) measures. Nahim et al. [1] appeared after the submission deadline. This survey lists as an advantage of the deterministic flow that "slight perturbations in the latent noise do not result in wildly different action sequences." The work it cites for this does not test the property, whereas our measurement does.

  2. It is credited to the action heads of the very policies we test. ACG [2] describes "generative flow-based VLA models (e.g., GR00T-N1, π0, SmolVLA), which offer improved stability and expressiveness." DiffusionVLA [3] credits "the robustness of high-frequency action generation offered by diffusion models." We already cite RobustVLA [4]. It attributes π0's "stronger intrinsic robustness" over its autoregressive variant to "the diffusion-based action head." Psiris et al. [5] write that π0.5 "casts action generation as flow matching for stable continuous control."

  3. It inherits a peer-reviewed inference about ODE flows that was never tested for flow-matching policies. Yan et al. [6] argued that neural ODEs are "intrinsically robust" because integral curves cannot cross. Kang et al. [7] record the view as established ("Recent studies have demonstrated that neural ODEs are intrinsically more robust") and then show that for classifiers the property has to be enforced. A flow-matching sampler is such an ODE with the latent noise as its initial condition, so the inference carries over directly. No analogous test existed for a trained few-step policy sampler before ours.

  4. No theorem decides it for action data. It must be measured, and we are the first to do so on deployed policies. Contraction of the probability-flow ODE is proven only under strong log-concavity (Gao and Zhu [9]). Without it the dynamics are "not contractive, and hence small errors quickly accumulate and are magnified" (Chen et al. [8]). Multimodal action distributions are not log-concave, so no result decides the sign for a VLA in either direction. CDP [10] responds by enforcing contraction in training, because "contraction pulls nearby flows closer to enhance robustness." The deployed, unregularised policies in use today are therefore left with a belief on one side and no theorem on the other. Our measurement fills that gap.

[1] N. Z. Nahim, F. N. Khan, P. Ju. A Survey of Flow Matching in Reinforcement Learning. arXiv:2610.16275, 2026.

[2] M. Park et al. ACG: Action Coherence Guidance for Flow-based Vision-Language-Action Models. ICRA, 2026.

[3] J. Wen et al. DiffusionVLA: Scaling Robot Foundation Models via Unified Diffusion and Autoregression. ICML, 2025.

[4] J. Guo et al. On Robustness of Vision-Language-Action Model against Multi-Modal Perturbations. ICLR, 2026.

[5] A. Psiris et al. Foundation Models in Robotics: A Comprehensive Review of Methods, Models, Datasets, Challenges and Future Research Directions. arXiv:2604.15395, 2026.

[6] H. Yan, J. Du, V. Y. F. Tan, J. Feng. On Robustness of Neural Ordinary Differential Equations. arXiv:1910.05513, 2019.

[7] Q. Kang, Y. Song, Q. Ding, W. P. Tay. Stable Neural ODE with Lyapunov-Stable Equilibrium Points for Defending Against Adversarial Attacks. NeurIPS, 2021.

[8] S. Chen, S. Chewi, H. Lee, Y. Li, J. Lu, A. Salim. The Probability Flow ODE is Provably Fast. NeurIPS, 2023.

[9] X. Gao, L. Zhu. Convergence Analysis for General Probability Flow ODEs of Diffusion Models in Wasserstein Distances. AISTATS, 2025.

[10] A. Abyaneh et al. Contractive Diffusion Policies. ICLR, 2026.

AuthorsOpenReview ↗

Response to Reviewer pE56 (2 of 2)

Response to Reviewer pE56 (2 of 2).

This comment states what the evidence in comment 1 establishes, how the paper will state its premise, and the changes to the paper.

What the evidence establishes. The belief appears in peer-reviewed work that credits flow-matching action heads with robust and stable action generation, and one survey states it for the latent noise itself [1] to [5]. For the latent noise the belief amounts to a claim about contraction. A perturbation of the noise that does not produce a "wildly different" action is one the noise-to-action map does not amplify, which is what σ_max(Φ) < 1 states. Nothing settles the question, because no theorem decides the sign for multimodal action data [8, 9] and the methods that want contraction enforce it in training [10]. The deployed policies in use are therefore credited with a property that theory does not grant and that no one had measured. This is the position the paper tests, and the revision states it in these terms.

What the measurement shows. Two of the four deployed policies expand on every episode and two contract (Fig. 2a). Contraction is therefore a property of the action head rather than of flow matching. The sharpest case is the one RobustVLA names. RobustVLA attributes π0's robustness to its flow head. π0's flow expands on 80/80 episodes, yet π0 is the most observation-robust policy in our panel (Table 2). SmolVLA also expands and is the least robust. The two expanders span the full range of observation robustness, so the sign of the flow carries no information about robustness. Whatever makes π0 robust, it is not contraction of its generation flow. This is finding (b).

On CDP. The reviewer reads CDP correctly, and the revision cites it only for the role it plays. CDP is evidence that the field treats contraction as the mechanism of robustness ("contraction pulls nearby flows closer to enhance robustness"), not evidence that flow matching contracts by itself. Sec. 2 already draws this distinction, and the abstract and Sec. 1 will match it.

Changes to the paper (C3). The abstract sentence becomes "Flow-matching action heads are credited with robust and stable action generation. For the latent noise this credits the flow with contracting perturbations, a property that recent methods enforce in training but that has not been measured on deployed policies. We show that it is not a property of flow matching." Sec. 1 ¶1 is reworded in the same terms and cites [1] to [10]. ¶4 replaces "the common belief" with "this inherited assumption." One Sec. 1 citation that does not support its sentence is corrected. DSRL [11] and Dense-Jump [12] are added to Related Work. They give independent evidence that π0's output depends strongly on its noise and that the flow-matching velocity field loses Lipschitz continuity near the end of the flow. No experimental result changes.

Other updates. Prompted by Reviewer zMjU, we re-derived the per-step criterion with the integration direction explicit. We also verified the two-expand, two-contract result on the deployed samplers (960/960 draws). Both are in the general response. Neither affects the cross-architecture split, the controls, the validation, or the noise-versus-observation distinction the reviewer highlights.

This addresses the gap the reviewer identified between the cited evidence and the position challenged. If the reviewer agrees, we would be grateful if this were reflected in their assessment, and we welcome any further question.

[11] A. Wagenmaker et al. Steering Your Diffusion Policy with Latent Space Reinforcement Learning. arXiv:2506.15799, 2025.

[12] Z. Chen et al. Dense-Jump Flow Matching with Non-Uniform Time Scheduling for Robotic Policies. ICRA, 2026.

Thanks for the rebuttal, which has addressed the major concerns. I have no more questions.

AuthorsOpenReview ↗

Response to Reviewer pE56 (closing)

We thank the reviewer for confirming that the response addresses the concerns. The uploaded revised PDF states the premise as in comment 2 and cites the sources (C3). Since this was the one weakness raised in the review and it is now resolved, we would be grateful if the reviewer would consider whether the overall rating still reflects their view of the paper.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response to all reviewers and the Area Chair

General response to all reviewers and the Area Chair.

We thank the reviewers for their careful reviews. The reviews recognise the cross-architecture study, the end-to-end measurement of the deployed sampler and the distinction between noise sensitivity and observation robustness. In short, the central result is unchanged and now verified on the deployed samplers. This comment summarises the shared points and lists every change to the paper (C1 to C7). The per-reviewer threads carry the full evidence. New measurements use the paper's 80-episode panel per model with the same seeds and draws, and every change in the revised PDF is typeset in blue.

1. The central result holds on the deployed samplers themselves (new). The paper's witness is restricted to the active 7-DoF block (Sec. 3.1, App. D). We have now measured the full 32-dimensional deployed sampler on every draw of the panel. For the expanders the exact Jacobian-vector product along the witness direction lower-bounds the deployed gain, and for the contractors we computed the full deployed Jacobian.

ModelVerdictDraws on the witness's side of 1Full deployed gain
SmolVLAexpands160/160at least 1.33 on every draw
π0expands320/320at least 2.21 on every draw
π0.5contracts320/320within 1.1% of the witness
GR00Tcontracts160/1600.76 to 1.01 of the witness

The split of two expanders and two contractors is a property of the deployed noise-to-action maps, verified on 960/960 draws. Paths through padding carry at most 3.3% of the gain and never change a verdict. The witness is now described as a faithful proxy rather than exact (C1), and the same data complete Table 3 (C2).

2. The premise is stated in peer-reviewed work, and this paper is the first to test it. Sec. 2 already frames the premise as inherited from control and sampler theory, enforced by CDP and untested on deployed VLAs. The abstract and Sec. 1 compressed this to "intrinsically contracts." The revision states the belief as the literature does. Flow-matching action heads are credited with robust and stable action generation in DiffusionVLA (ICML 2025), RobustVLA (ICLR 2026) and ACG (ICRA 2026), the last naming GR00T-N1, π0 and SmolVLA. A 2026 survey (arXiv, after the submission deadline) holds that "slight perturbations in the latent noise do not result in wildly different action sequences." For the latent noise this is a claim about contraction of the noise-to-action map. No theorem decides it for multimodal action data (Chen et al., NeurIPS 2023), and no prior work had measured it on a deployed policy (C3).

3. Read in the sampler's integration direction, the per-step criterion supports the end-to-end measurement. Reviewer zMjU correctly observes that the per-step object must follow the sign of dt. Re-derived with the orientation explicit and re-evaluated on all 960 draws, the per-step bound certifies contraction for GR00T (160/160) and π0.5 (292/320), certifies none for the expanders, and is never falsely satisfied. The disagreement reported for π0.5 in Table 1 and insight (a) came from reading λ_max(sym J) on a dt < 0 sampler. The revision restates the methodological point with the end-to-end gain as the quantity to report (C4). The re-oriented Arm C penalty contracts SmolVLA and brings π0 to the boundary. Every contracting arm still fails the non-degeneracy gates, so the tradeoff null stands (C5).

4. Expansion is within-mode sensitivity, not switching between action modes. A pre-registered probe at 40 states per model finds no distinct action modes at any state for the expanders, yet every one of their 160 draws expands. The contractors expand locally only at a decision state. GR00T's contraction is scoped to frame 0 and the App. F corollary is revised (C6).

Changes to the paper (revised PDF, within 10 pages). C1 witness described as the active-block restriction of the exact Jacobian, deployed-sampler audit in App. D. C2 Table 3 completed, gate (iii) at ε ≤ 0.03. C3 abstract and Sec. 1 aligned with Sec. 2, premise sources cited, DSRL (2025) and Dense-Jump (ICRA 2026) added to Related Work. C4 oriented per-step criterion in Prop. 3, App. A, Table 1, Sec. 5.2, insight (a) and contribution 2. C5 re-oriented Arm C in App. F. C6 multimodality probe as a new appendix section (App. G). C7 orientation stated once in Sec. 3.2, quantities introduced in words before symbols, derivations moved to the appendix.

If the reviewers find their concerns addressed, we would be grateful if this were reflected in their assessments.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.