acceptodds
Under review as a conference paper at ICLR 2027

Operating-point locality of soft constraints in flow-matching generative models

Abstract

Few-step flow-matching generators such as MeanFlow expose two operating modes—a one-function-evaluation (1-NFE) map and a multi-step Euler sampler— and neither respects, by construction, the continuous symmetries of the physical point clouds they are increasingly used to generate. We study soft constraints (a symmetry-encouraging equivariance penalty, and a physics sum-rule residual) added to the training loss, and ask whether a constraint enforced during training governs the samples. It does, but only locally: a soft constraint lowers the tar- get violation only at the sampling operating point—the (t,r) pair, and even the step count K—at which it was enforced, and relaxes elsewhere. Constraining the 1-NFE endpoint makes one-step samples equivariant (16×on SO(3) molecules, ∼600×on SO(2) jets) while multi-step samples stay at the unconstrained level; constraining the sampler’s velocity step does the reverse. A per-step audit locates the cause in a state mismatch, and a diffusion control isolates it. Enforcing the constraint at both operating points makes the generator equivariant at both (81×at 1-step and 3.1×at 10-step on molecules; 96×and 2.6×on jets), an equal-budget ablation shows this is coverage rather than regularization strength, a relocation control moves the effect by moving the training step count, and marginalizing over K yields one model equivariant at every sampling budget. The same law governs a physics constraint, and we measure its boundary: under a non-compact Lorentz boost the fix holds at one step only. This is a mechanism paper; it makes no sample-fidelity claim.

Primary area:
generative models
Keywords:
flow matching, soft constraints, generative model

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_L2oxOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

I think it's a interesting question to study where we know flow matching shows it can be used on physics constraint generation and we know meanflow learns average velocity between timesteps in flow matching where first impression should be we can see it generalize to meanflow in multistep setting with constraints.

However, experiments on SPICE molecules and JetClass jets show that constraints applied to one-step outputs often do not carry over to multistep sampling, and vice versa. The authors connect this to where the constraint is enforced—in terms of time pairs, states, and step counts—and propose combining endpoint and sampler-step penalties. Additional physics and Lorentz-boost experiments show where this approach helps and where it falls short.

Critical strengths

The controls make the main finding more convincing. The authors do more than show that a constraint fails to transfer. Increasing the weight of a single penalty does not fix the other sampling mode, while combining penalties at the same total weight helps both. Changing the training step count also moves the region of low constraint error. These experiments give good evidence that where the constraint is applied matters, not just how strongly it is weighted.

The paper gives a useful fix and shows its limitations. Combining endpoint and sampler-step penalties improves both sampling modes on molecules and jets, and varying the training step count improves robustness across the tested budgets. I also appreciate that the authors report the incomplete transfer of physics residuals and the failure under Lorentz boosts rather than presenting the fix as universally successful.

Critical weaknesses

I’m still unsure how much of this comes from the particular MeanFlow setup. Why I say this is that I remember in paper Physics-Constrained Flow Matching, they say even for flow matching, we need to uses repeated corrections of predicted final samples during inference to enforce physical constraints. With MeanFlow supporting both one-step and multistep generation, I would expect keeping soft constraints consistent across these settings to need extra care. But how much of the observed behavior comes from the MeanFlow objective, and how much reflects the broader problem of constraints not carrying over to sampling?

The practical benefit needs stronger sample-quality checks. I agree that a mechanism paper does not need state-of-the-art generation quality. Still, I would like to know whether the combined and variable-step-count penalties improve constraint satisfaction without substantially hurting quality or diversity. This matters because the physics residuals improve transverse momentum while worsening the jet-mass distribution. For those physics experiments, comparing against an applicable inference-time correction baseline would also clarify what the training-time remedy offers beyond correcting samples during generation.

The study uses Euclidean MeanFlow with added soft penalties. How should the locality findings be interpreted for geometry-aware formulations, such as in Riemannian Flow Matching paper and in Riemannian MeanFlow paper, where some constraints are built into the probability paths and sampling maps?

Questions for authors

How broadly should we interpret the claim that the recipe applies to any few-step generator? In particular, would you expect the same behavior for additional soft constraints in geometry-aware formulations?

What if with the projection step included for meanflow as what Physics-Constrained Flow Matching done for flow matching in meanflow multistep sampling case? How does the proposed fix compare with inference-time correction alone in constraint error, sample quality/diversity, and sampling time? Does training with the penalty reduce the amount of correction needed?

3 comments
AuthorsOpenReview ↗

Responses to Reviewer L2ox (part-1)

We sincerely thank the reviewer for the encouraging assessment and for the insightful questions on generality, sample quality, and scope. The numbers below refer to the revised PDF unless marked "submitted". The appendices are relettered: submitted A–E are now B, C, D, E, and G, with a new App. A (experimental details) and a new App. F (additional experiments). Due to the size limit, we have to truncate our replies into two parts.

The comments from the reviewer are quoted in boldface to help with navigation. We will italicize our highlighted points inlined in our replies without boldfacing them to avoid confusion.

"how much of the observed behavior comes from the MeanFlow objective, and how much reflects the broader problem of constraints not carrying over to sampling?"

We thank the reviewer for this question. Two pieces of evidence show that the effect is not specific to MeanFlow, one already in the submission and one new. First, DDIM is a single-mode diffusion sampler, and there the same penalty enforced at all tt retains a clear multi-step benefit, with the residual erosion matching the measured exposure bias (§4.2, App. B). Second, we retrained plain flow matching models on SPICE (App. F.2) with the same network and rr pinned to tt, so that the network learns only the instantaneous velocity, and swept KK up to 50. The results reproduce the MeanFlow pattern: the endpoint penalty improves the single-step sampler, and the interpolant-state velocity penalty lowers the violation more than twofold at every KK from 2 to 50. A shortcut model (App. F.5) shows the same dependence, with its self-consistency objective sharing part of a reduction across coupled budgets.

The dependence on the state distribution therefore holds for soft penalties on flow matching, shortcut, and diffusion samplers alike. MeanFlow adds the time-pair axis on top of it, which makes the choice of operating point a critical design decision and motivates our focus on two-time samplers. The sampler-agnostic statement is given in App. F.4 (Proposition 1): for any sampler built from Euler-type steps, the equivariance defect of the sample is the sum of the per-step velocity defects at the visited states, each carried through the remaining steps.

"The practical benefit needs stronger sample-quality checks … comparing against an applicable inference-time correction baseline"

We fully agree that this matters for practical use. We evaluated quality and diversity for the no-SEAL, endpoint, velocity, and combined models in both domains, with Chamfer and sliced-Wasserstein distances, per-feature marginals, jet observables, a short-range interatomic-distance diagnostic for SPICE, and a diversity statistic at fixed conditioning. At the deployed 10 steps, within-conditioning diversity is preserved within 10–20% in both domains, and on SPICE the constrained models match the unconstrained quality within seed noise. On jets, the combined penalty improves the per-particle marginals and the jet pTp_T and η\eta distributions, with a moderate cost in Chamfer distance and jet mass. The velocity-only penalty preserves single-step fidelity and is the most economical choice when only multi-step sampling is deployed. To stay within the page limit, the revised PDF summarizes these trade-offs in §6, and we will include the full tables in the appendix of the final version.

In addition, the new experiment on the released EPiC-FM generator (§4.5, Fig. 4, App. F.6) demonstrates the recipe on a model of published quality. A constraint on the high-level jet observable τ32\tau_{32} brings it close to the real-data level at the targeted step count, up to 61× closer than without the constraint, while the published sampler is preserved. The inference-time correction is discussed in our answer to the last question below.

"How broadly should we interpret the claim that the recipe applies to any few-step generator? … How should the locality findings be interpreted for geometry-aware formulations (Riemannian Flow Matching, Riemannian MeanFlow)?"

We thank the reviewer for raising the question of scope. Our finding has two parts, and we believe they both are broad enough.

The time-pair statement applies to any few-step generator that exposes a (t,r)(t, r) pair, including MeanFlow, flow maps, and consistency-type models, and §1 and §6 now state this scope in place of the broader submitted wording. The state-distribution statement is the more general result, since it requires only an iterative sampler and a penalty sampled from some distribution of states, and the DDIM experiment and App. F.2 confirm it beyond MeanFlow. Together, these statements yield a simple principle for any soft constraint added to few-step generators: a penalty acts on the states it is sampled from, so it should be placed where the sampler travels. This makes the result relevant wherever symmetries or physical constraints are imposed through training rather than architecture.

AuthorsOpenReview ↗

Responses to Review L2ox Part 2

Continuing on "How broadly should we interpret the claim that the recipe applies to any few-step generator? … How should the locality findings be interpreted for geometry-aware formulations (Riemannian Flow Matching, Riemannian MeanFlow)?"

Further on the architecture choice aspect, Geometry-aware models (Riemannian flow matching and the two Riemannian MeanFlow papers, now cited) build the manifold into the probability path and the sampling map, which is complementary to our analysis of additional soft terms. For soft terms added on top of such models, the distributional argument carries over directly, since every such term is still an expectation over states on the manifold. The Euclidean commutation argument holds in special cases, such as an isometric group action with an exponential-map update, and the revised limitations paragraph presents this extension as a well-motivated expectation that we plan to test. Preserving the manifold exactly does not by itself guarantee other symmetries or physical constraints, which is precisely where additional soft penalties. Therefore, we believe our analysis of them is useful.

"What if with the projection step included … How does the proposed fix compare with inference-time correction alone in constraint error, sample quality/diversity, and sampling time? Does training with the penalty reduce the amount of correction needed?"

We thank the reviewer for this valuable suggestion and have run this comparison. The baseline is a one-shot projection: each generated jet is rescaled once, at the final state, onto the conditioning targets (EE, pTp_T, mSDm_{SD}) on which the residual is trained, a deliberately simpler version of the repeated corrections of PCFM. We measured the residual after projection, the fraction of jets for which the closed form is clipped or infeasible, the wall-clock time of sampling plus projection, and the relative correction the projection applies to each model.

The projection meets the pTp_T sum rule exactly at negligible cost. Crucially, training with the residual places the samples close to the constraint surface, so the projection needs four to eight times less correction at the trained step count. The two approaches are therefore complementary, and residual training followed by projection gives the best combination of constraint error and local structure. One subtlety is that the conditioning uses the soft-drop mass, whereas the evaluated jet mass is that of the summed four-vector, so meeting the conditioning target exactly need not minimize the jet-mass Wasserstein distance. We could include the full comparison in the appendix of the final version, if permitted, as the complete set of experiments get overly heavy to conclude within the rebuttal period...

I thank authors for your extra experiments and reply. The comparison with ordinary flow matching using the same backbone, along with the shortcut-model results, largely addresses my concern that the behavior might be specific to meanflow. The clarification about geometry-aware models and additional soft constraints also makes sense to me. Hence, I maintain my positive opinion about this paper.

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

This paper studies whether few-step generators such as MeanFlow respect the symmetries of the physical point clouds that they train on. It was claimed that soft equivariance penalty on the 1-NFE endpoint makes one-step samples strongly equivariant but leaves multi-step samples at the unconstrained level. The paper then shows that enforcing the constraint at both the endpoint and the sampler-matched velocity step makes the generator equivariant.

Critical strengths
  • The paper asks an interesting question regarding whether few-step generators like MeanFlow can respect the symmetries underlying the training samples and attempts to understand this
Critical weaknesses
  • The paper is not easy to follow. Some sentences are confusing and not clear. For instance, in the question "a training-time constraint is enforced at some operating point; which operating points does it govern at sample time?", what does it mean to ask "which operating points does it govern at sample time"? The generative time of sampling is different from the one sampled during training, and they are not related in general unless we are in the one-step generation setting that maps the base distribution directly to a target distribution. In the one-step setting, we only use the one-step sampler and it is unclear why we should even bother with using multiple-step samplers. There are also numerous such unclear statements throughout the paper.
  • The fact that soft constraint on few-step generators is "operating-point-local" is not surprising and the fix proposed is straightforward. It is unclear what we gain from the studies provided in the paper. It does not provide a concrete method to improve symmetry preservation while retaining competitive generative performance.
  • There are missing details on the baselines and datasets in the experimental parts. It is also not clear what are the soft constraints (a detailed formulation is needed) and how they are implemented.
Questions for authors
  • Would similar claims hold for flow matching models such as CFM and Stochastic Interpolants (not just MeanFlow)?
  • Are there any reasons why different number of random seeds are used for the results in Table 1-2?
4 comments
AuthorsOpenReview ↗

Responses to Reviewer dLyj Part 1

We sincerely thank the reviewer for the careful reading and for the candid comments, which helped us improve the clarity and the motivation of the paper. The numbers below refer to the revised PDF unless marked "submitted". The appendices are relettered: submitted A–E are now B, C, D, E, and G, with a new App. A (experimental details) and a new App. F (additional experiments).

The comments from the reviewer are quoted in boldface to help with navigation. We will italicize our highlighted points inlined in our replies without boldfacing them to avoid confusion.

"what does it mean to ask 'which operating points does it govern at sample time'?"

We agree that this sentence was ambiguous, and we have rewritten it (§1) with a definition, together with other unclear statements in the manuscript. We hope these modifications could improve the overall clarity of the presentation. An operating point is the (t,r)(t, r) pair at which the network uu is evaluated during generation, together with the step count KK of the sampler. The single-step sampler evaluates uu once at (1,0)(1, 0), whereas the KK-step Euler sampler evaluates it at (tk,tk+1)(t_k, t_{k+1}). A soft constraint is enforced on uu at particular (t,r)(t, r) pairs and latent states during training, and the question is which samplers inherit it at generation time. The results show that the sampler at the enforced operating point inherits the constraint, and this observation leads directly to a recipe that covers every deployed sampler.

"In the one-step setting … it is unclear why we should even bother with using multiple-step samplers."

Few-step generators are valuable precisely because the same network serves both modes. The single-step map gives the fastest possible surrogate, and MeanFlow and its later variants improve fidelity from one to two or more steps at a still small cost. This flexibility is essential for scientific surrogates, where the sampling budget is chosen according to the available compute (§1). In our setting, the single-step map at the strongest penalty is equivariant but shifts the energy scale (§4.1, App. G), so the multi-step sampler is what one deploys for fidelity. A constraint on the single-step map does not automatically carry over to this sampler, and our recipe gives one model with low violation at every evaluated budget (§4.4), retaining both the speed of few-step sampling and the fidelity of multi-step sampling.

"not surprising and the fix proposed is straightforward … does not provide a concrete method to improve symmetry preservation while retaining competitive generative performance"

We appreciate this comment and have strengthened both the motivation and the evidence. Our multi-step equivariance probe is essential when assessing the sampling step flexibility of fast surrogates, as explained above. With it, we show that a penalty can be driven very small on its training objective while the deployed sampler remains unconstrained.

Further, our work makes three contributions. (i) The intuition that a penalty imposed at Gaussian noise does not constrain the evolved states is made precise in the new Proposition 1 (App. F.4): the equivariance defect of a sample is exactly the sum of the velocity defects at the states visited by the sampler. We also demonstrate, with more generative model paradigms, the generalizability of the finding. (ii) The experiments establish the less obvious parts of the picture. The per-tt endpoint penalty acts mainly at a single step even though it spans all t∈[0,1]t \in [0, 1], because the (t,r)(t, r) pair must match. Doubling a single penalty does not reach the other operating point at equal budget; retraining at K=2K = 2 moves the low point (Fig. 3b), and marginalizing over KK removes the off-grid spikes without diluting the penalty (Fig. 3c). (iii) The resulting method is a concrete training recipe: the endpoint term plus a velocity term at interpolant states, with KK marginalized when several budgets are used. It requires no architectural change and lowers the violation by nearly two orders of magnitude at a single step and about threefold at 10 steps on SPICE, with similar gains on JetClass.

Our revision further includes a released jet generator of published quality, and shows the same principle brings the high-level observable τ32\tau_{32} close to the real-data level at the targeted step count (§4.5, Fig. 4).

"missing details on the baselines and datasets … not clear what are the soft constraints (a detailed formulation is needed)"

We thank the reviewer for this comment and have added the details. The submitted draft described the variants in prose only. The revised §2 gives the violation functional (Eq. 1), the total loss, each variant as a choice of map gg and state distribution μ\mu, and the physics residual with its targets. Furthermore, App. A gives the datasets, splits, feature layouts, conditioning, model sizes, optimizer and schedule, and the other experimental settings.

AuthorsOpenReview ↗

Responses to Reviewer dLyj Part 2

"Would similar claims hold for flow matching models such as CFM and Stochastic Interpolants?"

For flow matching, yes. We have added a plain flow matching comparison with rr pinned to tt (App. F.2). On SPICE, the endpoint penalty improves only the single-step map, the Gaussian-state velocity penalty helps at the smallest budgets, and the interpolant-state velocity penalty lowers the violation more than twofold at every KK from 2 to 50, the same dependence as in MeanFlow. A shortcut model (App. F.5) and the DDIM experiment (§4.2) confirm it for further objectives. For stochastic interpolants, a nonzero noise term changes the state distribution along the path, so the penalty should be sampled at those states. Our interpolant-state penalty is the deterministic member of this family, and we see the stochastic case as a natural extension.

"Are there any reasons why different number of random seeds are used for the results in Table 1-2?"

The number of seeds reflects checkpoint availability: every independently trained seed of a variant is reported (3–5), and no seed was excluded because of its result. SPICE no-SEAL has four seeds because one run was a byte-identical duplicate of another seed and was removed. The submitted §1 sentence now refers to the equivariance tables, and App. A states the seed policy.

Reviewer_dLyjOpenReview ↗

Thank you for the rebuttal

I thank the authors for their rebuttal and revision. The responses have partly addressed my concerns but I find it tough to keep track of the changes made in the main paper and appendix. Please highlight all the changes made with color (blue or red) so that we can easily track what has changed from the previous version.

Thanks a lot for such a prompt follow-up! And apologize for the huge rewriting.

We have updated our supplementary materials, expanding them from the appendix PDF only to a zip file, which includes the requested diffs (both main text and appendix, supplied in red for deletions from the submitted version and in blue for additions in the current version).

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

This paper investigates the extent to which soft equivariance constraints or physics residuals imposed on MeanFlow generators remain effective during actual sampling. It distinguishes the constraint target, the time pair (t,r)(t,r), and the state distribution, showing that constraints applied only to the endpoint substantially reduce one-step equivariance violations but barely transfer to multi-step sampling. Through per-step audits of Gaussian-state velocity constraints, interpolant-state and time-matched constraints, equal-weight-budget comparisons, and relocation controls that change the training step count, the paper analyzes the importance of operating-point coverage.

On SPICE and JetClass, combining endpoint and velocity penalties reduces both one-step and 10-step equivariance violations under compact rotation groups. Training with a mixture of step counts KK extends the improvements across the tested sampling budgets. Similar sampler dependence is observed for physics sum-rule constraints, although these introduce trade-offs with other observables. Under Lorentz boosts, even the combined constraint fails to resolve the multi-step issue. Rather than claiming competitive sample fidelity, the paper positions its contribution as an empirical investigation of the conditions under which soft constraints transfer to sampling.

Critical strengths
  • The controlled experiments distinguishing where constraints are enforced from their effects on actual sampling are convincing. On SPICE, endpoint SEAL with λ=2\lambda=2 reduces the one-step equivariance violation from 0.269 to 0.017, while barely changing the 10-step violation from 0.214 to 0.206. In contrast, the sampler-time-matched velocity constraint evaluated at interpolant states reduces the 10-step violation to 0.064. The controls showing that increasing the weight of a single constraint does not sufficiently improve the other operating point support the interpretation that operating-point coverage, rather than merely constraint strength, is important.

  • The paper presents both the successes and limitations of constraint transfer, providing a practically useful diagnostic approach. The compact-group experiments use independently trained seeds across two domains and measure equivariance violations of the final generated outputs using masked probes. The combined constraint and KK-marginalization provide directly applicable design directions for models that support multiple sampling budgets. At the same time, the paper explicitly reports the multi-step failure under Lorentz boosts and the degradation in jet-mass modeling caused by physics residuals, while distinguishing constraint transfer from sample fidelity as a separate objective. Characterizing the boundaries of successful transfer, including failure cases, is more informative than simply reporting performance improvements.

Critical weaknesses
  • The experiments support the importance of operating-point coverage, but the mechanistic interpretation is stronger than what the controls directly establish. Table 4 shows that a velocity constraint trained on Gaussian states exhibits larger violations at evolved sampler states. However, the successful velocity constraint in Table 2 is evaluated on training interpolants ztz_t, rather than on the sampler's actual rollout states zkz_k. The Gaussian-versus-interpolant comparisons provide useful evidence that the state distribution matters, but they do not establish that the successful fix matches the sampler's state distribution or eliminates the remaining mismatch. The DDIM control provides complementary evidence, but it uses a separate hypersphere dataset and reports within-setup improvement ratios, as the authors acknowledge. It therefore does not constitute a matched comparison that independently isolates state coverage from differences in the model and data setting. A per-step audit of the successful interpolant-trained model would help determine how closely the proposed explanation accounts for the observed improvement.

  • The checkpoint selection procedure is not sufficiently specified. Appendix C states that checkpoints are selected based on observable quality, but does not clearly specify the selection split, the sampling step count KK used for selection, or the checkpoint selection criterion for SPICE. Table 3 reports the jet-pTp_T Wasserstein distance, which is also described as a criterion for checkpoint selection, making the separation between selection and final evaluation important. The choice of KK during checkpoint selection could also affect the observed locality curves. It is therefore important to clarify whether the results are obtained by evaluating a single fixed checkpoint across all KK, or by selecting the best checkpoint separately for each KK.

Questions for authors
  • Could the authors provide a per-step audit of the successful interpolant-trained velocity or combined model, comparing equivariance violations on interpolant states and actual rollout states at identical (t,r)(t,r) pairs? Does the fix reduce the gap in constraint violations between these states, or does it improve final-output equivariance despite a substantial remaining gap?

  • What are the data split, metric, and sampling step count KK used for observable-based checkpoint selection, and what is the selection criterion for SPICE? Is the selection data separate from the final evaluation data, and is a single fixed checkpoint evaluated across all KK? Could the authors also report the number of seeds and variability for each row in Table 3?

  • Could the authors clarify whether the approximately 1.15× MeanFlow result used in the DDIM comparison comes from the early, unmasked, single-seed setting in Table 5, where 0.254/0.220≈1.150.254/0.220 \approx 1.15?

2 comments
AuthorsOpenReview ↗

Replies to Reviewer MsmP

We sincerely thank the reviewer for the positive assessment and for the precise and constructive questions. The numbers below refer to the revised PDF unless marked "submitted". The appendices are relettered: submitted A–E are now B, C, D, E, and G, with a new App. A (experimental details) and a new App. F (additional experiments).

The comments from the reviewer are quoted in boldface to help with navigation. We will italicize our highlighted points inlined in our replies without boldfacing them to avoid confusion.

"the successful velocity constraint in Table 2 is evaluated on training interpolants rather than on the sampler's actual rollout states … A per-step audit of the successful interpolant-trained model would help."

We thank the reviewer for this suggestion and have added exactly this analysis (App. F.1, Table 4). We also want to clarify that the Table 2 entry itself is unchanged, since it is a final-sample measurement with 10 steps. At each Euler step of the sampler, we measure the relative violation of u(⋅,tk,rk)u(\cdot, t_k, r_k) at three kinds of states: the state the sampler actually reached, the interpolant state zt=(1−t)x+tϵz_t = (1 - t)x + t\epsilon at the same tt, and a fresh Gaussian state. We analyze the velocity-interpolant, combined, and KK-marginal models, together with the Gaussian-state velocity model of the original analysis.

These results help to audit the models, as the reviewer suggests: For every variant and at every step, the rollout-state violation lies within 0.85–1.38× of the interpolant-state violation, so the interpolant distribution is a reliable proxy for the states the sampler visits. The state-matched models are equivariant along their own trajectories and markedly less so at fresh Gaussian states, where they were never trained, while the Gaussian-trained models show the mirror image. Their final-output equivariance at 10 steps follows the rollout-state violation. The state-matched penalty therefore improves the final output because it closes the state-wise gap.

We add Proposition 1 with a straightforward derivation detailed in the current Appendix: the equivariance defect of the final sample equals the sum, over sampler steps, of the velocity defect at the rollout state, each carried through the remaining steps. On the trained SPICE models (Table 7), this sum tracks the measured final violation closely at rollout and interpolant states (rank correlation 0.97) and much less at Gaussian states (0.58). Following the reviewer's comment, the revised §4.2 presents the result as consistent with distributional coverage, with the per-step analysis as its direct test. We also re-derived the original analysis of submitted Fig. 2a. The gap between Gaussian and sampler states is 7–12× from the third step on.

Therefore, the conclusion remains unchanged.

"The checkpoint selection procedure is not sufficiently specified."

We thank the reviewer for pointing this out. Every equivariance and physics number uses one fixed checkpoint per seed, the same file at every KK, and nothing is selected on the probe or on a reported observable. This checkpoint is the epoch saved as best by the validation value of the training objective, computed each epoch on held-out files (JetClass 115–119, SPICE PubChem sets 8–9). Observable-based selection applies only to the scaling study of App. D.

This procedure has been stated clearly in the App. A of the current revision, which includes the experiment details. It includes the handling of runs resumed after preemption. Table 3 reports the number of seeds and the spread for every row.

To fully separate selection from evaluation, we repeated every probe with the same checkpoints on splits used for neither training nor selection with both SPICE PubChem and JetClass (details in the App. F).

To sum up, All conclusions are unchanged, and the ratios are preserved.

"whether the approximately 1.15× MeanFlow result used in the DDIM comparison comes from the early, unmasked, single-seed setting in Table 5"

Yes, and we thank the reviewer for catching this. Indeed, the value quoted in the submitted §4.2 and App. A came from an early probe that was unmasked and used a single seed (submitted Table 5). On the masked, seed-averaged probe, the retention of the MeanFlow endpoint penalty at K=10K = 10 is 1.04× to 1.09×.

The revision reports these corrected values in App. B, removes the unmasked table, and presents the two retention ratios as a qualitative contrast within each setup.

Thank you for your rebuttal. The new per-step measurements show that for the interpolant-trained models, violations on training interpolants closely track those on actual rollouts. I’m staying at weak acceptance because the Gaussian-state comparison with the same unnormalized penalty still rests on a single Gaussian-trained checkpoint.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Overall response

We sincerely thank all three reviewers for their careful reading and thoughtful comments. We are encouraged that the reviewers found the question interesting (dLyj, L2ox), the controlled experiments convincing (MsmP, L2ox), and the combined and KK-marginal recipes directly useful for models that support several sampling budgets (MsmP, L2ox). We have uploaded a revised PDF within the page limit. The title and the central claim are unchanged. The appendices are relettered: submitted A–E are now B, C, D, E, and G, with a new App. A (experimental details) and a new App. F (additional experiments).

We want to re-emphasize the motivations more clearly: few-step generators are the natural choice for fast scientific surrogates, since one network provides both a single-step map and a multi-step sampler, and the sampling budget can be chosen at deployment. Their standard evaluation relies on binned histograms of selected observables, which compare marginal distributions but cannot reveal whether a generator preserves the governing symmetry. Our multi-step equivariance probe tests the symmetry directly on every sample and at every budget. With this probe, we show where a soft constraint holds at generation time, and we derive a simple recipe that lowers the violation at every evaluated budget with a single model.

Changes to the manuscript.

  1. Clarity (dLyj). §1 defines the operating point and restates the central question; §2 gives every soft constraint explicitly as a choice of map and state distribution (Eq. 1) together with the total loss, and App. A documents the datasets, models, training settings, seed policy, and checkpoint rule. The changes are substantial.
  2. Checkpoint selection (MsmP). Every equivariance and physics result evaluates one fixed checkpoint per seed at every step count, selected by the validation value of the training objective, and nothing is selected on the probe or on a reported observable. We presented these in the appendix in the submitted version, and we have clarified this rule further in the current appendix. We further include the probes on the non-chosen splits as ablations (App. F.3).
  3. Minor corrections from our own re-derivation. The per-step analysis of §4.2 is re-derived with a reproducible script that is going to be published; Table 3 now reports the number of seeds for every row, the DDIM comparison uses the masked, seed-averaged probe (App. B), all SPICE comparisons use the same penalty weight, and the claims are scoped to the evaluated budgets and groups. These changes, while we feel necessary to report, don't modify the major conclusions.
  4. Scope and related work (L2ox). The recipe is stated for few-step generators that expose a (t,r)(t, r) pair, the state-coverage statement is separated as the more general result, and Physics-Constrained Flow Matching, Riemannian flow matching and MeanFlow papers are now discussed.

New results.

  1. A per-step analysis of the successful state-matched models shows that their violation on the sampler trajectory closely tracks the violation at the training interpolants (App. F.1, MsmP).
  2. We have added a proposition to show the central mechanism: the defect of the final sample is the sum of the velocity defects at the visited states (App. F.4, Proposition 1).
  3. Plain flow matching (App. F.2) and a shortcut model (App. F.5) reproduce the dependence on the operating point beyond MeanFlow (dLyj, L2ox), proving the generalizability of our findings.
  4. On the released EPiC-FM jet generator, which is considered the state-of-the-art performance model, we include a constraint on the high-level observable τ32\tau_{32} that brings this observable close to the real-data level at the targeted step count, up to 61× closer than without the constraint, while preserving the published sampler (§4.5, Fig. 4, App. F.6, L2ox). This addition suggests that our findings will hold with the state-of-the-art performative models.

Apologies again for submitting a heavily revised manuscript at a late stage, but we feel it necessary to conduct all such efforts, as we sincerely hope to improve the presentation of our findings.

Following the discussion with the reviewers, we have provided the diffs of the revisions in the supplementary material.

We provided them with text formatted in blue as additions to the current version and in red as deletions from the submitted version, zipped with the appendix PDF we originally provided with this revision.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.