acceptodds
Under review as a conference paper at ICLR 2027

Manifold-Stable Flow Matching

Abstract

Flow matching (FM) learns generative dynamics through velocity regression. A common assumption in geometric FM variants is that the prior distribution for the data is, in fact, a distribution on the data manifold. However, this requires knowledge of the data manifold, which in many applications is not available. In the absence of such prior knowledge, a low regression error alone does not guarantee that the FM conforms to the true data manifold. Manifold adherence keeps generated samples within valid configurations and is empirically associated with better task performance. In this paper, we introduce manifold-stable flow matching (MSFM), which can start from an arbitrary ambient prior (i.e., a prior distribution that is not necessarily on the manifold). The method leverages tools from nonlinear dynamics, namely contraction theory, to learn tangential transport with prescribed normal contraction to the manifold. Contraction theory provides a theoretical guarantee of convergence to the data manifold within a desired time window (e.g., one second). The construction uses analytical projectors when the data manifold is known, and local affine proxies estimated by principal component analysis of data samples when the data manifold is not known. We derive a family of probability paths compatible with our contraction guarantees and separate the training loss into a learnable manifold tangential term and a manifold normal residual. We establish manifold invariance and transverse convergence under explicit projection and proxy-selection assumptions. An ellipse experiment attains a mean terminal off-manifold error of order (controlled by our set numerical tolerance). In our Push-T robotic experiments, MSFM raises the success rate from to . In the Robomimic Square task, the success rate increases from to , while the rotation error order decreases from to (controlled by our set numerical tolerance). These results show stronger geometric adherence and higher observed task performance, produced by learning the data manifold in generative AI tasks.

Primary area:
generative models
Keywords:
Flow matching, contraction theory, geometric deep learning, manifold adherence

Reviews

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

This paper designed a flow matching approach to generate manifold constrained data with application to robotics. This paper designed an attraction along the normal direction to push the distribution toward the manifold. The paper is interesting but lacks some theoretical justification.

Critical strengths

The approach is new and interesting.

Critical weaknesses

Theoretical soundness:

  • Around line 157, the author stated that MM is embedded in Euclidean space, and tubular neighborhood guarantees uniqueness of projection. Is this a fact, i.e., for any embedded manifold MM there exists such a tubular neighborhood? If yes, is there a proof or reference to it? If no, is there any additional condition on MM so that it is true?

  • line 290: "Suppose the solution of (4) remains in U": I think the paper should establish under what conditions the trajectory / solution would remain in U.

  • If during sampling process, there exists a pair (t,xt)(t, x_t) s.t. xt∉Ux_t \notin U, either because of Gaussian as prior so that it's outside the tubular neighborhood, or due to the solution of (4) fails to remain in U, what would happen? Furthermore, would this cause bad regularity of the vector field, or some instability?

  • If a Gaussian prior is used in flow matching, will this violate the tubular neighborhood condition?

  • In flow matching, the population / true vector field (i.e. the exact minimizer of (1) in the submitted paper) can be written as E[X1−X0∣Xt=x]E[X_1 - X_0 \mid X_t = x]. Is there any similar formula for the proposed approach? If yes, does it rely on the tubular neighborhood?

Presentation:

  • The paragraph (line 162) "Why conventional flow matching does not imply stability." is difficult to follow. I suggest to remove the shorthand VM(x)V_M(x) and use dist(x,M)2dist(x, M)^2 throughout, and emphasize what is distance shrink and why it matters. Also, I don't think adding a dot on top of VM(x)V_M(x) is a good notation. If it represent time derivative, I suggest to simply wirte time derivative. It's a bit unclear to me by saying "only the normal component of the velocity matters", "this rate", "this pointwise condition".

  • Equation 4: I think there should be an intutive explanation for it. α\alpha is not intuitive (e.g., what it is, why it is needed, how you set it in practice?). Also, I suggest change the notation rr to something else or just remove that shorthand. What I mean is that, when I first look at the equation it's unclear to me what the term αr\alpha r is, and whether rr is a normal vector; on the notation side perhaps nn stand for normal, or some ⊥\perp could make it clear.

  • What does "define" mean in line 189? Is it specifying rr by an initial value?

  • line 196: when "integrating this prescribed linear ODE" is needed? Will this be expensive?

  • Section 4.2 is difficult to follow.

Possibly related work:

Questions for authors

See weakness.

Some other questions:

  • I have concerns about how difficult it is to compute time derivative of QtQ_t.

  • The normal direction is by construction and only tangent direction require neural network approximation?

Remarks: I am not familiar with robotics, which appears as an application of the proposed approach. Therefore, my review is on the theoretical side of the proposed approach, not to assess or judge its practical performance or applicability. I will raise my score if the presentation can be improved and the theoretical soundness question can be addressed.

3 comments
AuthorsOpenReview ↗

Response to Reviewer aerv (1/3)

We thank the Reviewer for their critical reading and constructive criticism. Revisions are highlighted in blue; see our general comment for a summary. Labels follow the review: T = Theoretical soundness, P = Presentation, R = Related work, Q = Questions.

T1 (tubular neighborhood): Does every embedded MM have a tubular neighborhood with a unique projection? Proof, reference, or extra conditions?

Response. We thank the Reviewer for their comment. Yes, every smooth embedded submanifold of Euclidean space without boundary admits a sufficiently small tubular neighborhood in which the nearest-point projection is unique and smooth. See Theorem 6.24 and Problem 6-5 in Lee, J.M. (2012). Sard’s Theorem. In: Introduction to Smooth Manifolds. Graduate Texts in Mathematics, vol 218. Springer, New York, NY. We stated these assumptions explicitly and added this reference to the manuscript.

The width of this neighborhood may vary along the manifold. A neighborhood of fixed positive width, which the revised Theorem 1 uses, always exists when the manifold is compact, but compactness is not necessary (as in the case of our noncompact manifold R3×SO(3)×R\mathbb{R}^3\times \text{SO}(3)\times \mathbb{R} for Square actions, for which 2\sqrt{2} is the fixed positive width).

T2 (staying in U): Establish under what conditions the solution of (4) remains in U (line 290).

Response. Thank you for this comment. The revised Theorem 1 now proves this instead of assuming it. Suppose every point closer to M than some fixed width has a unique, smooth nearest point on M. Then every trajectory that starts closer than this width stays closer for all flow times, and hence stays in U. The reason is that the distance to M cannot increase along the MSFM field, so the trajectory can never reach this width. The argument uses only this distance estimate, so it does not need uniqueness of solutions. Under mild conditions on the network, namely continuity and at most linear growth, the theorem also shows that these solutions exist up to the terminal time. Such a width always exists when M is compact; compactness is not necessary, and we give a noncompact example after the theorem (which is used in our Robomimic Square task). The proof is in Appendix E.

T3 (leaving U): What happens if xt∉Ux_t \notin U, from a Gaussian prior or from leaving U? Does this cause poor regularity or instability?

Response. Thank you for this comment. By the revised Theorem 1, trajectories that start within the fixed width described in our preceding response remain there. The remaining issue is initialization farther from the manifold. We added a short explanation in the main text and a new Appendix A to specifically address this comment. We describe an optional step before MSFM sampling. First, choose an anchor for each sample and keep it fixed. Move the sample straight toward this anchor until it is closer to the anchor than the fixed width in Theorem 1. Then start MSFM sampling. With this step, any ambient prior can be used, including a Gaussian.

Appendix A states the required assumptions, explains why samples enter the region, and gives examples of suitable regions.

Continued in part 2 (reply below).

AuthorsOpenReview ↗

Response to Reviewer aerv (2/3)

T4 (Gaussian prior): Does a Gaussian prior violate the tubular-neighborhood condition?

Response. Thank you for this comment. A Gaussian prior has full support, so some samples can lie outside any fixed-width neighborhood of a curved manifold. For a fixed affine proxy, this is not an issue, since its projection is defined everywhere. Following the revisions addressing the preceding comments, we now explicitly describe how to extend the manifold-convergence guarantee to any ambient prior, including a Gaussian prior, under the conditions stated in Appendix A (a known safe width around the manifold, an anchor on the manifold for each sample, and the same conditions on the contraction rate and the network as in Theorem 1).

Samples already within the fixed width can initialize MSFM directly. For samples farther away, an outer stage contracts each sample toward its selected anchor until it is closer to the anchor than the fixed width in Theorem 1. MSFM then runs inside U, where invariance and prescribed normal contraction ensure convergence to the manifold. Appendix A provides the construction, assumptions, and entry criteria, and notes that training must then use the distribution of samples after the entry step. For the rotation manifold SO(3), we further show that the guarantee holds for almost every initial matrix, so a Gaussian prior needs no extra step there (Appendix B). In our revised Square experiment, the rotation source is not Gaussian: every source rotation block lies inside the 2\sqrt{2} tube, so Theorem 1 and Lemma 1 apply directly (Appendix H).

T5 (population field): Is there an analogue of E[X1−X0∣Xt=x]E[X_1 - X_0 \mid X_t = x] for MSFM, and does it rely on the tubular neighborhood?

Response. Thank you for this question. Yes. For square-integrable training velocities and a model class that can represent the optimum, the population-optimal MSFM output is P(x)E[ut∣xt=x]−α(t)r⊥(x),P(x)\mathbb E[u_t\mid x_t=x]-\alpha(t)r^\perp(x), where P(x)P(x) and r⊥(x)r^\perp(x) are determined by the current state. This follows directly from Proposition 1: regression learns the tangential conditional mean, while the normal component is prescribed. Here, utu_t is the training-path velocity; it equals X1−X0X_1-X_0 for linear paths.

Under normal compatibility, Q(xt)ut=−α(t)r⊥(xt)Q(x_t)u_t=-\alpha(t)r^\perp(x_t), this expression reduces to E[ut∣xt=x]\mathbb E[u_t\mid x_t=x], recovering the usual flow-matching marginal field. Recovering the target distribution additionally requires the learned field to have unique solutions, as in standard flow matching.

The conditional-mean identity itself does not require a tubular neighborhood. For a known curved manifold, the geometric quantities must be defined; smooth projection is guaranteed within U. For a fixed affine proxy, they are globally defined. Whether this field equals the usual flow-matching field does depend on the neighborhood. For a curved manifold, we prove this equality when the source lies within the fixed width of Theorem 1 (Lemma 1 in Appendix C). For SO(3), we prove it for almost every source matrix (Appendix B). We have added this characterization immediately after Proposition 1.

P1 (line 162 paragraph): Hard to follow; use dist(x,M)2\mathrm{dist}(x,M)^2 and an explicit time derivative instead of V˙M\dot V_M; clarify "this rate" and related phrases.

Response. Thank you for pointing this out. We have rewritten this paragraph using squared distance directly and an explicit time derivative. It now explains that tangential velocity makes no contribution to the distance derivative, while inward normal motion reduces the distance and corrects off-manifold deviations. We also state that FM training and inference do not control this rate, so the distance can grow at some sampling states.

P2 (Equation 4): Explain α\alpha intuitively and how it is set; clarify whether rr is a normal vector.

Response. We now call r⊥(x)=x−Π(x)r^\perp(x)=x-\Pi(x) the normal residual and explain that its length is the distance to the manifold. Equation 4 labels the tangential and normal terms with parallel and perpendicular notation. The text explains that the network learns tangential motion, while the prescribed normal term pulls samples toward the manifold. The rate α(t)\alpha(t) controls the strength of this attraction. We give α(t)=1/(1−t)\alpha(t)= 1/(1-t) as a practical example and point to Section 4.4 and the experimental appendices for the schedule and settings.

Continued in part 3 (reply below).

AuthorsOpenReview ↗

Response to Reviewer aerv (3/3)

P3 (line 189): What does "define" mean? Is rr specified by an initial value?

Response. Yes. The initial normal residual is fixed by the source sample: r0⊥=x0−Π(x0)r_{0}^\perp=x_0-\Pi(x_0). The ODE then determines its value along the chosen base curve. We replaced “define” with “Obtain the normal residual rt⊥{r_{t}^\perp} by integrating the ODE” and made this construction explicit in Section 4.1.

P4 (line 196): When is integrating the prescribed linear ODE needed, and is it expensive?

Response. This ODE is used to construct training pairs. Numerical integration is needed for a general curved geometry when its solution is not available in closed form. The affine proxies and the SO(3) construction used here have explicit solutions, so they require no additional numerical ODE solve for the training path. For other geometries, the cost depends on the geometry and numerical solver. We have clarified this in Section 4.1.

P5 (Section 4.2): Section 4.2 is difficult to follow.

Response. We have added short explanations of the local PCA proxy, its tangent and normal projectors, and the normal residual. The revised text describes how the field combines learned tangential motion with attraction to the selected proxy. Additionally, we added Figure 2 in Section 4.2 to visualize the mathematical terms used in constructing our MSFM formulation.

R1 (related work): arXiv 2510.22044 seems similar and is worth discussing.

Response. Thank you for this reference. We agree that OLLA shares the idea of tangential motion combined with prescribed normal correction. OLLA uses Langevin dynamics with a specified target density and constraint functions, and proves convergence to the constrained target distribution under its assumptions. MSFM learns transport from data samples using analytical projectors or local PCA proxies. We now cite Jeon et al. (2025) and discuss this relationship in Related Work.

Q1 (time derivative of QtQ_t): How difficult is it to compute the time derivative of QtQ_t?

Response. Thank you for raising this concern. Section 4.1 now clarifies that Q˙t\dot Q_t is needed only to construct training targets, not during inference. For a general smooth known manifold, only the product Q˙trt⊥\dot Q_t r_t^\perp is required. It can be evaluated analytically or by directional automatic differentiation along the base curve, holding the residual fixed, without forming the full projector Jacobian. Its cost depends on the geometry. For each fixed affine proxy, Q˙t=0\dot Q_t=0. For SO(3), the explicit path in Appendix B is differentiated directly to obtain the training velocity, avoiding explicit evaluation of Q˙t\dot Q_t and an additional numerical ODE solve.

Q2 (normal vs. tangent): Is the normal direction set by construction, with only the tangent direction learned by the network?

Response. Yes. The normal velocity −α(t)r⊥(x)-\alpha(t)r^\perp(x) is prescribed by the geometry and contraction schedule. The network affects only the tangential velocity through P(x)wθ(t,x)P(x)w_\theta(t,x). Proposition 1 shows that only the tangential loss depends on the network parameters.

Reviewer_HEUUOpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes manifold-stable flow matching for generating samples near a data manifold. It combines learned tangential motion with normal contraction, using known geometry or local PCA estimates. The authors prove convergence to the known manifold or estimated proxies, and show better geometric accuracy and task success in robitics experiments.

Critical strengths
  1. The connection between the algorithm and geometric guarantee is clear. MSFM separates learned tangential motion from prescribed normal contraction, and accounts for changing normal directions on curved manifold. For known manifolds, the theory gives explicit decay rates for the distance to the manifold under stated assumptions. This is useful because a small flow-matching loss alone doesn't guarantee that generated samples remain close to the manifold.

  2. The results on robotic tasks are promising. The reported success rate improved from 74% to 82% on Push-T and from 60%-72% on Square, together with better geometric accuracy. In Square, both methods use the same final rotation projection, which makes the comparison more convincing. These results suggest that controlling geometry during generation can improve task performance.

Critical weaknesses
  1. The unknown-manifold theory stops at the PCA proxies. Theorems 2–3 prove convergence to estimated affine spaces or their union, but the paper does not bound their distance from the true manifold or explain how PCA errors affect convergence. An affine space can extend far from a curved manifold, so approaching the proxy does not guarantee approaching the true geometry. This limits the theoretical support for a central setting of the method.

  2. The experimental evidence for reliable performance gains is limited. The robotic experiments use 50 rollouts per task, but do not report variation across independent training runs. The success gains correspond to four additional successful episodes on Push-T and six on Square. These results are promising, but leave uncertainty about whether the improvements are consistent across training runs. Since improved task performance is an important part of the paper’s contribution, this weakens the empirical case for acceptance.

Questions for authors
  1. For unknown manifolds, can you bound the distance to the true manifold using the PCA approximation error? What assumptions are needed to connect the proxy convergence results to the true manifold?

  2. Are the success gains on Push-T and Square consistent across independent training runs? Please clarify the number of training runs and report the variation in success rates, if available.

3 comments
AuthorsOpenReview ↗

Response to Reviewer HEUU (1/2)

We thank the Reviewer for their critical reading and constructive criticism. Revisions are highlighted in blue; see our general comment for a summary.

W1 (proxy vs. true manifold): Theorems 2–3 stop at the PCA proxies; bound the distance to the true manifold and the effect of PCA errors.

Response. Thank you for raising this point. Our contribution is a contraction framework that guarantees convergence toward the geometry used by the model. The PCA construction demonstrates how this framework extends beyond explicitly known manifolds to geometry estimated from data.

We chose local PCA as a simple, established method for estimating tangent spaces, drawing on Zhang and Zha (2004), Singer and Wu (2012), and Kaslovsky and Meyer (2014). Our use of local proxies is inspired by the manifold-adherence calculation in Pan et al. (2026), “Much Ado About Noising: Dispelling the Myths of Generative Robotic Control.” Developing a new manifold estimator is outside the scope of this work.

Theorem 2 guarantees convergence to a fixed proxy, and Theorem 3 to the union of proxies when the closest proxy is selected at each state, as flow time approaches the prescribed terminal time, t=1t=1. Accuracy relative to the true unknown manifold consequently depends on how well this geometry approximates it. We have added a bound after Theorem 3 with three terms: the normal residual to the proxy, tangent-estimation error measured by the chordal distance between the estimated and true tangent spaces times the displacement along the proxy from its anchor, and a curvature term quadratic in that displacement. The bound assumes a C2C^2 manifold, anchors on the manifold, the correct tangent dimension, and that the projected points remain in valid local patches. This locality condition is essential because an affine proxy extends beyond the region where it approximates a curved manifold. Appendix E now gives the derivation and defines the local curvature constant.

Other geometry estimators could be incorporated by constructing the required projections and normal dynamics and verifying the corresponding assumptions. Comparing these estimators and extending the analysis to their representations are left for future work.

W2 (variation across runs): Gains rest on 50 rollouts per task; report variation across independent training runs.

Response. Thank you for raising this concern. We have expanded the Push-T experiments to measure variation across independent training runs. We now use 30 paired training seeds and evaluate each trained model on 50 episodes with repeated block-position disturbances. This gives 1,500 episodes per method. The data split and evaluation scenarios remain fixed across training seeds.

MSFM increases mean success from 69.1% to 78.5%. The 95% confidence intervals are [63.4%, 74.5%] for FM and [74.4%, 82.3%] for MSFM. The mean gain per seed is 9.3 percentage points, with a 95% confidence interval of [1.3, 17.4]. These intervals describe uncertainty in mean success across training runs for the fixed evaluation scenarios.

MSFM also increases mean maximum reward from 0.820 to 0.871 and reduces mean off-proxy error from 0.304 to 5.74×10−55.74 \times 10^{-5}. The revised Push-T figure shows the mean training and validation losses and their standard deviations across seeds, together with confidence intervals for success. Appendix G gives the full evaluation details.

We have also re-run the Square experiment with 20 paired training seeds. FM and MSFM are evaluated after the same fixed number of training updates, on the same 50 initial states, giving 1,000 episodes per method.

MSFM increases mean success from 58.2% to 68.5%. The 95% confidence intervals are [55.2%, 60.8%] for FM and [66.0%, 71.0%] for MSFM. The mean gain per seed is 10.3 percentage points, with a 95% confidence interval of [6.6, 14.0]. MSFM also reduces the mean distance of generated rotations from SO(3), before the final projection, from 7.4×10−27.4 \times 10^{-2} to 5.6×10−75.6 \times 10^{-7}. Figure 5 shows the training and validation losses across seeds, together with confidence intervals for success and cumulative success. Appendix H gives the full evaluation details.

Q1 and Q2 are answered in part 2 (reply below).

AuthorsOpenReview ↗

Response to Reviewer HEUU (2/2)

Q1 (bound to the true manifold): Can the PCA approximation error bound the distance to the true manifold, and under what assumptions?

Response. Thank you for this question. We have added a bound after Theorem 3, with its derivation in Appendix E. Let Pi⋆P_i^\star be the true tangent projector at xi∗x_i^\ast. In a valid local patch, we obtain dist⁡(x,M)≤∥ri⊥(x)∥+∥Pi(x−xi∗)∥∥Pi−Pi⋆∥F2+Ci∥Pi(x−xi∗)∥2.\operatorname{dist}(x,\mathcal M)\leq \lVert r_i^\perp(x)\rVert + \lVert P_i(x-x_i^\ast)\rVert \frac{\lVert P_i-P_i^\star\rVert_F}{\sqrt{2}} + C_i\lVert P_i(x-x_i^\ast)\rVert^2. The three terms are the distance to the proxy, the tangent-estimation error, and the curvature error. The bound assumes a C2C^2 manifold, xi∗∈Mx_i^\ast\in\mathcal M, and rank⁡(Pi)=dim⁡M\operatorname{rank}(P_i)=\dim\mathcal M. Here ∥Pi−Pi⋆∥F/2\lVert P_i-P_i^\star\rVert_F/\sqrt{2} is the chordal distance between the estimated and true tangent spaces, and CiC_i bounds the local second-order departure from the true tangent plane. Appendix E provides the precise patch condition and the definition of CiC_i.

Connecting this to Theorem 2 (fixed proxy) or Theorem 3 (closest-proxy selection) requires uniform bounds on the displacements along the proxies from their anchors, chordal distances, and curvature constants within the valid visited patches. Under these conditions, the true-manifold distance is bounded by a decaying normal-error term plus a geometric error floor. The contraction theorems alone do not establish these additional conditions.

The true tangent projector and curvature constant are generally unknown, so this decomposition explains the sources of geometric error without making the bound directly computable from the PCA proxies alone. Deriving statistical bounds for PCA estimation and establishing the required locality conditions remain beyond the present analysis.

Q2 (consistency across runs): Are the gains consistent across training runs? Report the number of runs and the variation.

Response. Thank you for this question. Please see our response to your second critical weakness. There, we report the updated Push-T results from 30 paired training seeds and Square results from 20 paired training seeds, with 50 episodes per seed and method, and give the mean success rates and 95% confidence intervals across training seeds.

Thank you for the detailed rebuttal and additional experiments. Both my concerns are fully addressed. I am happy to raise my score.

Reviewer_nVRUOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper introduces a modification to flow matching that aims to guarantee landing on a target manifold or constraint set. The method, MSFM, explicitly decomposes the generative flow into a predetermined normal flow and a learned tangential flow, so that a suitable choice of normal flow will ensure reaching the manifold. This can be applied to manifolds known in closed form, or with learned data geometries, for which the authors use knn-based local PCA. The problem is well motivated, although the exposition could be made slightly clearer. There are theoretical guarantees provided, given assumptions on the normal flow. The method is evaluated experimentally, although I think the scope and scale of these experiments needs more exploration, and the dependence on the method on nice manifold geometries is not sufficiently explored. I also feel that, since the method interferes significantly with standard flow matching, it needs more explicit positioning to fully inform practitioners on where using this would be appropriate.

The paper is an interesting algorithmic contribution, I would be willing to increase my score if my questions are addressed.

Critical strengths
  1. The problem is well motivated and explained, with strong, clear writing and diagrams.
  2. The theoretical guarantees are clear and useful.
Critical weaknesses
  1. The exposition of the two methods in 4.1 and 4.2 is heavy on notation and would benefit a lot from a diagram.
  2. The intrinsic geodesic solution depends on several geometric factors: as a minimum the space must be connected, and the efficiency of the intrinsic path as a generative trajectory will depend a lot on curvature, with highly curved spaces leading to much longer paths than ambient-space flow matching. I think these topics need to be more fully discussed and ablated in experiments. For example, if I fix a manifold dimension and progressively increase its curvature, how do your results vary? Similarly, what happens if the space is disconnected? On the other hand, these curvature issues may reduce in high dimensions, so additionally varying intrinsic and/or ambient dimension would also be useful information. The current synthetic example is just an ellipse in 2d, and I think the method requires more full evaluation.
  3. The data-estimated geometry uses local PCA for the tangent spaces, which is unstable under noise - it would be good to consider or at least discuss robust kernel-weighted methods e.g. [1,2]. Since your motivation is high dimensional data like robotics imagery, you should evaluate with or discuss neural approaches to metric learning [3], or use the Jacobian of the score [4].

[1] Singer, A., & Wu, H. T. (2012). Vector diffusion maps and the connection Laplacian.

[2] Jones, I. (2026). Manifold diffusion geometry: Curvature, tangent spaces, and dimension.

[3] Bamberger, J., et al. "Riemannian Metric Matching for Scalable Geometric Modeling of Distributions."

[4] Kharitenko, A., Shen, Z., De Santi, R., He, N., & Dorfler, F. (2026, April). Landing with the score: Riemannian optimization through denoising.

Questions for authors
  1. You discuss in the conclusion how FM and MSFM are evaluated at small scale where MSFM does better. How do you expect this performance to vary with scale? I would expect that, given enough data and scale, a method that intervenes in the flow matching pipeline so noticeably will struggle to compete in the end. However, there are many low-data settings where this would not be applicable. Where do you see the method sitting wrt data and model scale?
2 comments
AuthorsOpenReview ↗

Response to Reviewer nVRU (1/2)

We thank the Reviewer for their critical reading and constructive criticism. Revisions are highlighted in blue; see our general comment for a summary.

W1 (exposition of Sections 4.1 & 4.2): Add a diagram for the notation-heavy Sections 4.1 and 4.2.

Response. Thank you for this suggestion. We added a diagram (Figure 2 in Section 4.2) that complements Figure 1 and illustrates the constructions in Sections 4.1 and 4.2. It shows a sampling trajectory with its shrinking normal residual, the roles of the tangent and normal projectors PP and QQ, the learned tangent component and prescribed normal feedback, and the extension to local PCA affine proxies. The diagram also shows the tangential training loss and normal-compatibility condition. Its caption explains each panel, and Section 4 refers to it.

W2 (geometry ablations): Test curvature, dimension, and disconnected support beyond the 2D ellipse.

Response. We added controlled experiments to assess how curvature, intrinsic dimension, ambient dimension, and disconnected support affect manifold adherence, generated path lengths, and final-distribution matching.

We vary curvature using ellipses with a fixed perimeter. We vary intrinsic dimension while fixing ambient dimension, and ambient dimension while fixing intrinsic dimension. We also test two disjoint circles, with supplied component weights and training pairs formed within each component. FM and MSFM use matched training and sampling budgets across multiple training seeds. We measure terminal distance to the known manifold and generated path lengths. We also measure maximum mean discrepancy (MMD) between raw generated endpoints and independent target samples.

MSFM maintains much smaller terminal manifold distances throughout the tested settings. It also has lower mean MMD across the curvature and ambient-dimension sweeps and at almost all intrinsic dimensions. FM has very slightly lower mean MMD at only one of the intrinsic dimensions. For disconnected support, MSFM improves both manifold adherence and mean MMD. These results show that the geometric advantage persists across the tested settings, and is almost always accompanied by better distribution matching.

Mean MSFM trajectories are shorter in every tested curvature and dimension setting, but about 9% longer for disconnected support.

Path length and computational cost remain distinct. Even a longer path does not mean more computational cost, since regardless of the path, the main computational cost is from the ODE solver steps for the transport from prior (time 0) to the target (time 1). However, MSFM may incur higher inference time because its field requires additional geometric operations, particularly when projection is computed iteratively.

Appendix I provides the complete experimental details and results.

Our Robomimic Square task also tests MSFM beyond simple synthetic geometry. The model generates action horizons in an 832-dimensional ambient space. The target manifold has 448 intrinsic dimensions and 384 normal directions. It contains 64 curved rotation manifolds, each described by SO(3). Their tangent and normal spaces change as the rotations change. Square therefore tests high-dimensional generation with many nonlinear geometric constraints. The model must handle these constraints while learning coordinated translation, rotation, and gripper commands for closed-loop robot control.

W3 (tangent estimation): Local PCA is noise-unstable; discuss robust kernel methods [1,2] and, for high-dim data, learned metrics [3] or score Jacobians [4].

Response. Thank you for these suggestions. Our local PCA construction is inspired by the manifold-adherence calculation in Pan et al. (2026), “Much Ado About Noising: Dispelling the Myths of Generative Robotic Control,” which measures the component of a predicted action outside the subspace spanned by nearby expert actions.

Estimating manifold geometry from data is a distinct problem from establishing stability toward a chosen geometry. Here, we use a simple PCA-based construction to demonstrate that our geometric guarantees also apply to data-estimated geometry, beyond explicitly known manifolds. Under the stated assumptions, these guarantees concern the estimated proxies; they do not remove geometry-estimation error or establish robustness to noisy data.

The revised manuscript now discusses kernel-weighted local PCA and diffusion geometry [1,2], neural Riemannian metric matching [3], and geometry estimation from denoising scores and their Jacobians [4]. These are relevant alternatives for learning manifold geometry. Comparing these estimators and extending our guarantees to each representation are outside the scope of the present paper and are left for future work. We have also clarified the motivation and scope of our PCA construction in Section 4.2 and Related Work.

Q1 is answered in part 2 (reply below).

AuthorsOpenReview ↗

Response to Reviewer nVRU (2/2)

Q1 (scaling): How will MSFM's advantage change with data and model scale, given that it intervenes noticeably in flow matching?

Response. Thank you for this question. An important distinction is between restricting the ambient velocity field and restricting the distributions that can be represented on the target manifold.

MSFM prescribes normal contraction while retaining learned transport in every tangential direction. Since the tangent projector acts as the identity on tangent vectors, a sufficiently expressive network can represent the required tangential field after projection. Conventional FM learns normal and tangential motion together through velocity regression; MSFM explicitly separates these roles.

Proposition 1 and the discussion after it make this distinction precise. For normal-compatible training paths, the population-optimal MSFM field equals the conventional FM marginal field for those paths. Under the stated geometric and transport assumptions, the target distribution can therefore still be recovered. Prescribed normal contraction does not inherently require sacrificing distributional expressiveness. We have added a short explanation immediately after Proposition 1.

Our current experiments show improvements at matched, compact capacity. As noted in the Conclusion, MSFM changes only the output field, so it can be combined with larger backbones and longer training. Whether its geometric advantages persist at state-of-the-art scale is left for future work.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of revisions

We thank the reviewers for their constructive feedback. In the revised PDF, changes addressing the reviews are highlighted in blue, and text condensed for the 10-page limit is highlighted in magenta. Main changes:

  • Theory (aerv): Theorem 1 now proves that trajectories starting within a fixed distance of the manifold stay in the tubular neighborhood and exist up to the terminal time (Appendix E). Appendix A covers priors with samples outside this neighborhood, such as a Gaussian. Appendix B shows that for SO(3) the guarantee holds for almost every initial matrix. The population-optimal field is characterized after Proposition 1.
  • Unknown manifolds (HEUU): a new bound after Theorem 3 relates the distance to the true manifold to the proxy distance, the tangent-estimation error, and curvature (Appendix E).
  • Presentation (aerv, nVRU): rewritten Sections 3.1 and 4.1, a new Figure 2, and an expanded Related Work.
  • Experiments (HEUU, nVRU): Push-T with 30 paired training seeds: success 69.1% → 78.5% (mean gain 9.3 points; Appendix G). Square re-run with 20 paired seeds: 58.2% → 68.5% (mean gain 10.3 points; Appendix H). New Appendix I tests curvature, intrinsic and ambient dimension, and disconnected support.
  • Supplementary code: the Square code now matches the 20-seed re-run, with trained models for one seed, episode records and loss histories for all 20 seeds, and scripts that reproduce the reported Square statistics and figures; the Ellipse and Push-T code is unchanged.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.