acceptodds
Under review as a conference paper at ICLR 2027

Rethinking Riemannian MeanFlow: Geometric Gaps and Geometry-Aligned Flow Maps

20 views

Abstract

MeanFlow enables efficient one-step or few-step generation by learning average velocities over finite time intervals. Recent works have extended this principle to Riemannian manifolds for data with intrinsic non-Euclidean structure, such as rotations and directions. However, under curvature, the Euclidean relationships among average velocity, endpoint displacement, and generation error no longer directly hold. We revisit two formulations of Riemannian MeanFlow and identify two geometric gaps. First, parallel-transported path averages generally differ from endpoint log-displacements, potentially inducing one-step endpoint errors and flow-map composition defects. Second, even with endpoint-based representations, metric distortion through the exponential map causes tangent-space residuals and geodesic endpoint errors to weight directions differently. Motivated by these findings, we propose **Barycentric Semigroup Riemannian MeanFlow (RMF-BSC)**, which combines endpoint differential consistency with multi-split geodesic semigroup consistency to align training with endpoint geometry. We establish barycentric stability bounds and a variance-reduction result for multi-split averaging. Experiments validate the identified gaps and show that RMF-BSC achieves strong one-step performance across four diverse manifold benchmarks.

Primary area:
generative models
Keywords:
Riemannian MeanFlow, Manifold Generative Modeling, One-Step Generation, Endpoint Geometry, Semigroup Consistency

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

This paper analyzes the errors in two existing Riemannian MeanFlow methods and identifies three types of errors: endpoint errors, composition errors, and tangent-space errors. The authors provide detailed mathematical analyses of these errors, along with relatively clear proofs. To address these issues, the paper proposes an approach based on endpoint representations and multi-split geodesic semigroup consistency. The experiments first verify the existence of these errors, supporting the theoretical analysis. In subsequent experiments on real-world datasets, the proposed method achieves the best performance on several benchmarks across SO(3), Earth, RNA, and DNA.

Critical strengths

The paper identifies and addresses error issues in existing Riemannian MeanFlow methods. The proposed method is relatively robust to hyperparameter choices, and the additional computational cost remains within an acceptable range.

Critical weaknesses
  1. The experimental results show that the proposed method does not achieve the best performance on all datasets. Does this suggest that its effectiveness still depends on the specific dataset, rather than providing a general solution to the identified errors? For example, if the model makes accurate predictions, there should be no tangent-space errors. Does the BSC loss still work if the manifold is not a Hadamard manifold?

  2. Potential errors in JVP computation. In practice, JVP is used to compute the covariant derivative and covariant differentiation. However, JVP only provides the differential term and does not directly include the Christoffel symbol terms. Could this introduce additional hidden errors? If so, how much would these errors affect the actual results?

Questions for authors

see weaknesses.

4 comments

We sincerely thank the reviewer for the thoughtful and constructive review and for the positive assessment of our paper’s soundness and informativeness. We particularly appreciate your recognition of our analysis of the errors in existing Riemannian MeanFlow formulations and of the experiments supporting that analysis. We are also grateful for your acknowledgment of the method’s robustness to hyperparameter choices and its acceptable computational overhead. It is encouraging that the paper helped deepen your understanding of these geometric issues. Your questions about dataset-dependent performance, applicability beyond Hadamard manifolds, and the implementation of covariant derivatives are valuable for clarifying both the scope of our claims and the numerical realization of the method. We address each point below and will revise the manuscript to make these aspects clearer.

Q1: Dataset-dependent gains, exact predictions, and applicability beyond Hadamard manifolds.

A: We thank the reviewer for these questions. We agree that the gains from BSC depend on the data distribution, manifold geometry, and model fit. Our aim is to identify and address geometric mismatches in finite-time generation, rather than guarantee the best performance on every dataset. In particular, the effect of the metric gap depends on displacement length and error direction; when this effect is weak or predictions are already sufficiently accurate, the additional benefit may be limited.

Regarding accurate predictions, we distinguish the two gaps. For the representation gap, perfectly fitting the parallel-transported path-average velocity does not necessarily recover the endpoint log-displacement required by the sampler. Thus, zero regression error need not imply zero endpoint error. For the metric gap, we agree with the reviewer: when the endpoint prediction is exact, both tangent-space and geodesic losses vanish. Our claim concerns how these losses weight nonzero prediction errors, rather than an incorrect realizable fixed point of endpoint-based methods. This distinction is discussed in Appendix D.1.

BSC also applies to non-Hadamard manifolds, as illustrated by our sphere and SO(3) experiments. The Hadamard assumption ensures a globally unique barycenter and the associated stability inequality; it is not required to define the BSC loss. Appendix E.2 provides a local extension: when the relevant endpoints lie in a common geodesically convex region and the Fréchet objective satisfies suitable strong-convexity conditions, a local stability guarantee holds with a curvature- and region-dependent constant. Uniqueness and smoothness can fail near the cut locus, so we do not claim an unconditional global guarantee on arbitrary non-Hadamard manifolds.

Q2: JVP connection corrections and their numerical effects.

A: We thank the reviewer for raising this concern. We agree that a standard JVP provides only the differential term and does not automatically include the connection correction. For Eulerian/LDI, however, the relevant quantity is the combination of both derivative terms in the complete training target. In the left-trivialized representation of SO(3), their connection corrections are 12v×ϕ\tfrac12v\times\phi and 12v×u\tfrac12v\times u, respectively. Since ϕ=Logx(y)=−hu\phi=Log_x(y)=-hu within the chosen logarithm branch, their contributions cancel: Tcov−Traw=−12v×ϕ−h2v×u=0.T_{\mathrm{cov}}-T_{\mathrm{raw}} =-\frac12v\times\phi-\frac h2v\times u=0. Thus, an individual JVP differing from a covariant derivative does not imply an error in the combined LDI target.

To quantify the practical effect, we froze the model weights and inputs on Swiss (SO(3)) and Earthquake (S2S^2), and compared the targets, losses, and gradients of all trainable parameters computed with raw JVPs versus full geometric corrections. On the sphere, the correction uses tangent-space projection. For Eulerian and BSC, all three discrepancies remain at numerical precision, approximately 10−15−10−1410^{-15}-10^{-14}. These results show no observable target, loss, or gradient error from using raw JVPs for Eulerian/LDI or BSC in the tested settings.

This conclusion does not directly extend to PT, whose target lacks the log-map derivative term needed for the cancellation. For PT on Swiss, the target, loss, and relative gradient discrepancies are 0.1860.186, 0.04290.0429, and 0.02790.0279, respectively. On Earthquake, the target and loss discrepancies are 0.4720.472 and 0.06870.0687, while the gradient discrepancy remains near numerical precision. The effect of connection terms therefore depends on the complete objective and geometric representation.

Table Target, loss, and gradient discrepancies between raw-JVP and covariant computations.

DatasetMethodΔT\Delta_TΔL\Delta_LΔgrel\Delta_g^{\mathrm{rel}}
SwissPT1.86×10−11.86\times10^{-1}4.29×10−24.29\times10^{-2}2.79×10−22.79\times10^{-2}
SwissEulerian9.17×10−169.17\times10^{-16}1.78×10−151.78\times10^{-15}1.76×10−161.76\times10^{-16}
SwissBSC7.11×10−157.11\times10^{-15}1.78×10−151.78\times10^{-15}2.18×10−162.18\times10^{-16}
EarthquakePT4.72×10−14.72\times10^{-1}6.87×10−26.87\times10^{-2}1.94×10−151.94\times10^{-15}
EarthquakeEulerian9.74×10−169.74\times10^{-16}3.47×10−183.47\times10^{-18}2.55×10−152.55\times10^{-15}
EarthquakeBSC6.11×10−166.11\times10^{-16}2.60×10−182.60\times10^{-18}7.13×10−167.13\times10^{-16}

Thank you for the detailed response and additional experiments. The authors have addressed all of my concerns.

Thank you for your time and constructive feedback, which has helped us improve the paper. We are glad that our responses and additional experiments have addressed all of your concerns.

Reviewer_h25vOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper revisits Riemannian MeanFlow through two geometric distinctions: transported path-average velocities need not coincide with endpoint log-displacements, and tangent-space residuals need not measure geodesic endpoint errors faithfully. Building on endpoint-based RMF, the authors propose RMF-BSC, combining log-displacement differential supervision with multi-split geodesic semigroup consistency. The paper provides geometric counterexamples, local error characterizations, and barycentric and variance-reduction arguments, followed by experiments on rotations, Earth-event locations, RNA torsion angles, and conditional DNA generation.

Critical strengths
  1. Useful conceptual clarification. The distinction between path averages and endpoint displacements is relevant to one-step sampling on curved manifolds. The paper also correctly separates conditional interpolation paths from marginal-flow trajectories, avoiding the mistaken inference that geodesic conditional paths imply geodesic marginal trajectories. The explicit spherical counterexample makes the representation mismatch concrete.

  2. Targeted geometric diagnostics and a meaningful residual ablation. The controlled experiments demonstrate that endpoint and composition errors can arise independently of neural approximation. Importantly, RMF-Tan matches RMF-BSC in the number of splits, differential objective, normalization, and mixture weight, isolating the effect of replacing tangent residuals with geodesic residuals. This provides more direct evidence than an overall-method comparison alone.

  3. Promising empirical results with reported uncertainty. The main experiments report results over five seeds and cover several geometric settings. Improvements are particularly encouraging on Swiss Roll and on the reported DNA metrics. The appendix also provides training configurations, numerical safeguards, and per-step timing, while explicitly acknowledging several limitations of the theoretical arguments.

Critical weaknesses
  1. The incremental contribution over existing methods is insufficiently isolated. Endpoint RMF already provides log-displacement supervision, while GFM already uses geodesic semigroup consistency. RMF-BSC mainly combines these components with multi-split averaging, which reduces conditional gradient variance but does not change the expected objective. A controlled comparison with the corresponding single-split formulation is needed to establish the significance of this combination.

  2. The connection between the geometric gaps and generation quality remains incomplete. The diagnostics primarily use prescribed trajectories rather than learned benchmark flows. Endpoint or composition errors do not necessarily imply worse generated distributions, and the metric gap changes error weighting rather than the realizable fixed point. Measuring these gaps along learned trajectories would better support the explanation for the observed gains.

  3. Training efficiency and baseline comparisons require stronger controls. The default K=4K=4 increases per-step training time by approximately 31-44% over K=1K=1, but the paper does not establish an advantage under matched training time. Comparisons should also consider the strong G-LSD variant and clarify whether baseline parameterizations and stabilization settings follow their recommended configurations.

  4. The biological metrics do not fully support the claimed capabilities. RNA MMD measures individual angle marginals and cannot detect incorrect joint dependencies; pairwise visualizations provide only partial support. For DNA, pooled FBD and 6-mer correlation do not directly measure agreement with the specified transcription profiles. Joint-distribution and condition-alignment metrics are needed to substantiate these results.

Questions for authors
  1. Under matched architecture, time sampling, anchoring, normalization, and training time, how much benefit remains from adding multiple splits to the existing endpoint/geodesic consistency formulation?

  2. How large are the representation and metric gaps along learned benchmark trajectories, and do they predict which datasets or samples benefit from RMF-BSC?

  3. Can the RNA results be evaluated using a metric sensitive to the full joint distribution, and can DNA generation be evaluated for agreement with the specified conditioning profiles?

  4. How do the results compare with G-LSD and the recommended parameterization and stabilization choices of Endpoint RMF? Please distinguish controlled objective comparisons from comparisons against each method's strongest practical configuration.

4 comments

We thank the reviewer for the careful reading and constructive feedback. We appreciate your recognition of the conceptual clarification between path averages and endpoint displacements, including the distinction between conditional interpolation paths and marginal-flow trajectories. We also thank you for highlighting the controlled geometric diagnostics and the matched RMF-Tan ablation, which isolates the effect of geodesic residuals. We are encouraged by your positive assessment of the five-seed empirical results, particularly on Swiss Roll and DNA, and your acknowledgment of the implementation details and stated theoretical limitations. Below, we address your concerns point by point.

Q1: Benefits of multiple splits under matched training time.

A: We thank the reviewer for the suggestion. Our primary goal is to accelerate sampling through one- or few-step generation, rather than reduce training cost. We acknowledge that multiple splits increase the computational cost per training update, making it important to assess whether their benefits persist under the same training budget. We therefore added controlled comparisons between single- and multi-split formulations, matching the architecture, time sampling, anchoring, normalization, and total training time to evaluate the contribution of multiple splits under an equal computational budget.

The table below shows that increasing KK from 1 to 4 yields no consistent additional benefit under matched training time: Cone improves marginally (0.0122→0.01210.0122 \to 0.0121), while Fisher and Volcano favor K=1K=1. Both configurations nevertheless outperform all listed baselines in mean MMD, indicating that BSC’s advantage is already present with a single split. Thus, although multi-split averaging reduces split-induced conditional gradient variance per update, this does not necessarily translate into better generation under a fixed training-time budget.

Table Fair-training results on three datasets with different split K.

MethodConeFisherVolcano
RFM0.1752 ±\pm 0.00100.0723 ±\pm 0.00060.3650 ±\pm 0.0133
RCT0.0310 ±\pm 0.00380.0278 ±\pm 0.00070.2020 ±\pm 0.0359
G-PSD0.0341 ±\pm 0.00330.0306 ±\pm 0.00060.1851 ±\pm 0.0371
G-LSD0.0304 ±\pm 0.00380.0234 ±\pm 0.00080.1100 ±\pm 0.0391
RMF-PT0.0295 ±\pm 0.00360.0263 ±\pm 0.00070.1888 ±\pm 0.0447
RMF-Eulerian0.0293 ±\pm 0.00360.0265 ±\pm 0.00060.1881 ±\pm 0.0460
RMF-Semi0.0300 ±\pm 0.00360.0234 ±\pm 0.00080.1916 ±\pm 0.0427
BSC K=1K=10.0122±0.0012‾\underline{0.0122 \pm 0.0012}0.0209 ±\pm 0.00060.0731 ±\pm 0.0122
BSC K=4K=40.0121 ±\pm 0.00130.0212±0.0005‾\underline{0.0212 \pm 0.0005}0.0781±0.0089‾\underline{0.0781 \pm 0.0089}

Q2: Geometric gaps along learned trajectories and their relationship to performance.

A: Thank you for this question. Since the ground-truth marginal flow trajectories underlying the benchmark data are unknown, we cannot directly measure the representation and metric gaps along these trajectories. Instead, we quantify both gaps along numerically integrated learned flows, extending our diagnostics beyond prescribed trajectories.

The measurements show substantial dataset dependence. For example, along RMF-BSC’s learned trajectories, the representation gap is approximately 2.78×10−82.78\times10^{-8} on Line and 0.03010.0301 on Swiss, while the actual-residual-direction metric distortion, ∣Qlocal−1∣\lvert Q_{\mathrm{local}}-1\rvert, is approximately 0.0028%0.0028\% and 16.65%16.65\%, respectively. These values are averages over ten training seeds.

The generation benchmarks provide complementary evidence. The observed improvements of Eulerian RMF over PT on several datasets support the practical relevance of endpoint-based representations, while the comparison between RMF-BSC and RMF-Tan in Table 4 supports using geodesic endpoint distances for consistency training. Swiss exhibits relatively large geometric discrepancies together with strong RMF-BSC generation performance, which is consistent with our proposed explanation.

We emphasize that RMF-BSC does not require these geometric discrepancies themselves to vanish. Its design uses endpoint log-displacements and measures consistency directly in endpoint geometry. In particular, the metric discrepancy affects finite-error weighting rather than the realizable zero-error solution. The learned-flow diagnostics and generation results therefore support the practical relevance of this design.

Table Metric gap and Test MMD on Fisher and Swiss.

DatasetMethodMetric gap (%)Test MMD ↓\downarrow
FisherRMF-PT0.2974±0.02360.2974 \pm 0.02360.027835±0.0018900.027835 \pm 0.001890 (↑0.56%\uparrow 0.56\%)
FisherRMF-Eulerian0.2974±0.02360.2974 \pm 0.02360.027832±0.0018890.027832 \pm 0.001889 (↑0.54%\uparrow 0.54\%)
FisherRMF-BSC0.3007±0.02390.3007 \pm 0.02390.027681±0.001844\mathbf{0.027681 \pm 0.001844}
SwissRMF-PT15.838±0.82615.838 \pm 0.8260.095092±0.0025960.095092 \pm 0.002596 (↑134.24%\uparrow 134.24\%)
SwissRMF-Eulerian14.020±0.71514.020 \pm 0.7150.073711±0.0025230.073711 \pm 0.002523 (↑81.57%\uparrow 81.57\%)
SwissRMF-BSC16.552±0.71016.552 \pm 0.7100.040596±0.002731\mathbf{0.040596 \pm 0.002731}

Q3: RNA joint-distribution evaluation and DNA conditioning consistency.

A: We thank the reviewer for the suggestion. For RNA, we agree that NLL provides an evaluation sensitive to the full joint distribution and has been used in prior work. However, its computation requires multi-step ODE integration to evaluate the model density, which is distinct from the one-step sampling performance emphasized in our paper. We therefore did not report it in the original manuscript. Following your suggestion, we have added RNA NLL evaluations, and our method continues to perform well on this joint-distribution metric.

Table Comparison results on RNA with test nll for full joint distribution.

MethodTest NLL ↓\downarrow
RFM-4.9371
G-PSD-4.9550
G-LSD-4.6408
RMF-PT-3.5323
RMF-Eulerian-3.5323
RMF-BSC-4.9968

For DNA, following prior RMF works[1-2], we additionally report paired Sei-MSE between the Sei-predicted signals of generated sequences and their corresponding reference sequences under the same conditioning in  Table 1. This provides a proxy measure of condition-related functional agreement, complementing FBD and 6-mer correlation, which assess overall distributional agreement and sequence statistics. We have added these results to Section 6.3 (pages 8–9). All corresponding changes are highlighted in red.

Table Comparison of our method and baselines in the MSE metric.

MethodMSE ↓\downarrow
RFM0.0296±0.00040.0296 \pm 0.0004
RCT0.0648±0.00020.0648 \pm 0.0002
G-PSD0.0443±0.00030.0443 \pm 0.0003
RMF-PT0.0544±0.00030.0544 \pm 0.0003
RMF-Eulerian0.0332±0.00040.0332 \pm 0.0004
RMF-Semi0.0306±0.00050.0306 \pm 0.0005
RMF-BSC0.0283±0.0004\mathbf{0.0283 \pm 0.0004}

[1] Woo D, Skreta M, Park S, et al. Riemannian MeanFlow[J]. ICML, 2026.

[2] Zhong Z, Sun H, Zhao Y, et al. Riemannian MeanFlow for One-Step Generation on Manifolds[J]. ICML, 2026.

Q4: G-LSD comparisons and controlled versus best-tuned Endpoint RMF configurations.

A: We thank the reviewer for the suggestion. We have added experimental results for G-LSD in several benchmarks (Cone, Swiss Roll and Volcano). RMF-BSC continues to achieve the best performance across these benchmarks.

We also clarify that the baselines in the original manuscript already use their respective best tuned hyperparameters, while sharing the network architecture and other common experimental settings to ensure fair comparisons. Thus, our comparisons allow method-specific hyperparameter tuning rather than imposing identical hyperparameters across all methods.

Table Comparison with existing baselines and the added G-LSD baseline.

MethodConeSwiss RollVolcano
RFM0.14960.12580.1708
RCT0.06470.12350.1452
G-PSD0.04430.08450.1310
G-LSD0.0348$\underline{0.0625} $0.1311
RMF-PT0.02760.12230.1248
RMF-Eulerian$\underline{0.0275} $0.08280.1397
RMF-Semi0.03050.0833$\underline{0.1146} $
RMF-BSC0.02160.05820.1117

Thank you for the detailed rebuttal and additional experiments, which clarify several points; however, given that multi-split averaging shows no consistent benefit under matched training time, the contribution beyond Endpoint RMF and GFM's geodesic semigroup consistency still appears limited to me, so I will keep my current score for now.

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

The paper revisits two existing Riemannian MeanFlow formulations: (1) RMF-PT (Zhong et al.) defines the average velocity by parallel-transporting velocities along the trajectory, and (2) Endpoint RMF (Woo et al.) uses the endpoint log-map. It identifies two “geometric gaps”: the path–endpoint representation gap and tangent-endpoint metric gap. The proposed solution, RMF-BSC, combines a log-displacement-identity (LDI) loss with a semigroup consistency loss. The semigroup loss is measured as squared geodesic distance and averaged over K random split times. Extensive experimentation on four benchmarks evaluated with three metrics indicates superiority.

Critical strengths

Strengths:

  1. Clear and correct identification of the representation gap.
  2. The authors have included the limits of their own results, such as the composition bound being one-sided and multi-split averaging being an estimator change
  3. Theoretical proofs are detailed, with all assumptions made explicit.
  4. A complete ablation study that covers the role of lambda, endpoint geometry, and computational overhead.
Critical weaknesses

The novelty over RMF seems incremental.

The authors have deviated from standard evaluation metrics such as MSE.

Identical values in two tables raise concerns about baseline runs. It would be great if the authors could test their method for negative curvatures.

Questions for authors

Questions:

  1. LDI looks exactly like the Eulerian RMF objective. This paper’s Eq. (11) is the same target written in the backward time convention. The RMF paper also already uses an epsilon-floored interval weighting and adaptive loss weighting, both of which appear here. So RMF-BSC is precisely RMF-Eulerian plus a multi-split geodesic semigroup term. Please clarify the novelty or differentiating factor from RMF.
  2. The RMF paper uses a different metric (MSE). Have the authors compared the baselines on this metric as well? Also, please justify your choice of an alternate metric.
  3. Interestingly, RMF-PT and RMF-Eulerian are the same up to 4 decimal places in both Table1 and Table2. Please justify.
  4. Every curved benchmark considers positive curvature. Have the authors explored negative curvature, similar to "Generalised Flow Maps for Few-Step Generative Modelling on Riemannian Manifolds", which explores the Poincaré disk?
6 comments

We thank the reviewer for the careful reading and constructive feedback. We appreciate your recognition of our clear and correct identification of the representation gap, our explicit acknowledgment of the limitations of the theoretical results, and the detailed proofs with clearly stated assumptions. We also thank you for highlighting the comprehensive ablation study covering loss weighting, endpoint geometry, and computational overhead. Below, we address your concerns point by point and clarify the relationship to prior work, the experimental findings, and the scope of our contributions.

Q1: Relationship to Eulerian RMF and novelty of BSC.

A: Thank you for raising this point. We agree that u^LDI=−∇v~t(1)log⁡xt(x^rθ)−hDtCuθ(xt,r,t)\widehat{u}_{LDI} = -\nabla_{\widetilde{v}_t}^{(1)} \log_{x_t}(\widehat{x}_r^{\theta}) - h D_t^C u_{\theta}(x_t,r,t) is the Eulerian RMF target expressed in our backward-time convention. Epsilon-flooring and adaptive loss weighting also follow existing RMF practice and are not contributions of our work. We will make this attribution explicit. This equivalence concerns the LDI component; the BSC component uses a different error measure from RMF’s tangent-space regression objectives.

Specifically, Semigroup RMF compares the logarithmic displacements of the direct and composed predictions in a common tangent space, whereas BSC compares the resulting endpoints directly through their geodesic distance on the manifold. Writing these endpoints as ydirecty_{\mathrm{direct}} and ycompy_{\mathrm{comp}}, the respective discrepancies, before time weighting, are ∥log⁡x(ydirect)−log⁡x(ycomp)∥gx2\lVert \log_x(y_{\mathrm{direct}}) - \log_x(y_{\mathrm{comp}}) \rVert _{g_x}^{2} and dg2(ydirect,ycomp)d_g^{2}(y_{\mathrm{direct}},y_{\mathrm{comp}}).

These quantities are generally inequivalent on curved manifolds. Thus, BSC changes the geometry in which finite composition errors are penalized, beyond averaging over multiple splits. Our pullback-metric analysis characterizes this distinction and explains its relevance to endpoint prediction. The reported advantage of RMF-BSC over RMF-Tan in Table 4 provides empirical support for this design.

Our contribution therefore combines geometric analysis with endpoint-based multi-split training. We characterize the representation gap between transported path averages and endpoint log-displacements, which Endpoint RMF already avoids, and further analyze the remaining tangent–endpoint metric discrepancy. Motivated by this analysis, RMF-BSC measures composition discrepancies directly through geodesic distances between manifold endpoints and aggregates multiple intermediate-time splits. We establish a Fréchet-barycentric interpretation and conditional gradient-variance properties for this formulation. The reported improvement over RMF-Tan in Table 4 provides empirical support for the endpoint-based error measure. Thus, while the LDI component follows Eulerian RMF, the differentiating contributions lie in the geometric characterization, the intrinsic multi-split consistency formulation, and its theoretical and empirical evaluation.

Q2: DNA evaluation metrics and Sei-MSE comparisons.

A: Thank you for the suggestion. We recognize the value of Sei-MSE used in RMF: it measures the discrepancy between the Sei functional predictions for generated DNA sequences and their paired reference sequences, providing a useful assessment of functional consistency.

Our choice of FBD and 6-mer correlation was motivated by evaluating DNA generation quality from the perspectives of aggregate distributions and sequence statistics. Specifically, FBD measures distributional differences between generated and reference sequences in the Sei feature space, while 6-mer correlation assesses agreement in local sequence-pattern statistics. This choice reflects our focus on the quality of the generated distribution and is consistent with our use of MMD on the other datasets. Sei-MSE complements these metrics by providing an additional comparison of paired functional predictions.

In response to the suggestion, we add an additional MSE comparison, which shows that RMF-BSC achieves the lowest error among the seven evaluated methods, with 0.0283±0.00040.0283 \pm 0.0004, compared with 0.0332±0.00040.0332 \pm 0.0004 for RMF-Eulerian / LDI, and also outperforms RFM and RMF-Semi. These results demonstrate that RMF-BSC is competitive under the functional prediction metric adopted by RMF. MSE, FBD, and 6-mer correlation provide complementary perspectives on generation quality by assessing paired functional predictions, feature distributions, and local sequence statistics, respectively. We have added the Sei-MSE results to Table 2 and Section 6.3 (page 8), along with a description of the metric in Section 6.1. We have also updated the corresponding material in Appendix F.2 (page 24) and Appendix F.5.2 (page 27). All changes are highlighted in red.

Table Comparison of our method and baselines in the MSE metric.

MethodMSE ↓\downarrow
RFM0.0296±0.00040.0296 \pm 0.0004
RCT0.0648±0.00020.0648 \pm 0.0002
G-PSD0.0443±0.00030.0443 \pm 0.0003
RMF-PT0.0544±0.00030.0544 \pm 0.0003
RMF-Eulerian0.0332±0.00040.0332 \pm 0.0004
RMF-Semi0.0306±0.00050.0306 \pm 0.0005
RMF-BSC0.0283±0.0004\mathbf{0.0283 \pm 0.0004}

Q3: Matching RMF-PT and RMF-Eulerian results.

A: Thank you for pointing out this observation. The matching values in Tables 1 and 2 require separate explanations. For the RNA experiment in Table 2, the data lie on the flat torus T7\mathbb{T}^{7}. In local angular coordinates away from the logarithm’s branch cuts, −∇vt(1)log⁡xt(y)=vt.-\nabla_{v_t}^{(1)}\log_{x_t}(y)=v_t. Consequently, the Eulerian RMF target reduces to vt−hDtuθv_t-hD_tu_\theta, coinciding with the RMF-PT target used in this experiment. Our implementation also applies the same validity mask to both objectives. Their matching results under identical training and evaluation settings therefore have a mathematical basis. We will clarify this special case in the revision and avoid presenting these entries as independent evidence for distinct objectives.

For the Fisher and Line datasets on SO(3) in Table 1 the archived evaluations we inspected contain nonidentical MMD values, with absolute differences approximately between 10−910^{-9} and 10−810^{-8}, which are indistinguishable at four decimal places. Fisher is a rotationally symmetric, 24-mode distribution, whereas Line concentrates probability mass along three curves. These multidirectional structures may lead to cancellation among directional training signals. In the archived results, both methods produce small rotational displacements on these datasets. In this regime, the curvature-dependent difference between RMF-PT and RMF-Eulerian is second order in the displacement, resulting in nearly identical generated samples and MMD values that coincide to four decimal places.

Q4: Evaluation on negatively curved manifolds.

A: Thank you for the suggestion. To address this, we conducted an additional generation experiment on the Poincaré disk with sectional curvature −1-1, learning a wrapped-normal target distribution. All methods use the same network architecture, data splits, and number of training updates.

In this evaluation on the Poincaré disk, BSC achieves an average MMD of 0.0101, lower than the other methods with valid results reported in the table. These results demonstrate BSC’s generation capability on a negatively curved manifold, supporting its applicability beyond positively curved settings. We have added these results to Section 6.3 (pages 8–9). All corresponding changes are highlighted in red.

Table Comparison of our method and baselines on Poincaré disk.

MethodMMD ↓\downarrow
RFM0.1809±0.00030.1809 \pm 0.0003
RCT0.0218±0.00490.0218 \pm 0.0049
G-PSD0.0152±0.00310.0152 \pm 0.0031
RMF-PT0.0272±0.00140.0272 \pm 0.0014
RMF-Eulerian0.0103±0.00240.0103 \pm 0.0024
RMF-Semi0.0106±0.00240.0106 \pm 0.0024
RMF-BSC0.0101±0.0022\mathbf{0.0101 \pm 0.0022}

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.