acceptodds
Under review as a conference paper at ICLR 2027

Validity-Aware Flow Matching on Lie Groups: What Transfers to Target-Free Generators?

Abstract

Flow matching learns a marginal vector field from conditional training paths, yet many notions of validity concern the trajectory itself. Recent methods improve those conditional paths through geometry-aware or cost-aware interpolation, or impose constraints during fitting and inference. None of them settles which path properties survive marginalisation into a target-free generator. We study that gap on compact Lie groups, accounting for population marginalisation, finite-capacity regression and discrete sampling. At the population optimum, conditional flow matching preserves every one-time marginal, including terminal validity under state-only criteria, and expected additive state-time costs. The path law is another matter. Conditional paths and the population flow they induce can agree on every one time marginal and still differ in region hitting, path length and endpoint coupling. Integrated cost alone cannot control endpoint violations. Control becomes available once the deployed path satisfies a temporal modulus, of the kind a speed cap or a pathwise kinetic-energy bound supplies. We trace these effects on tori, rotation groups and molecular conformations using bounded, endpoint-pinned Lie-group paths with exact left-trivialised velocities. The experiments separate endpoint coupling, path design, regression and numerical sampling, and show where a conditional-path improvement survives deployment and where a validity claim needs evidence from the deployed generator itself.

Primary area:
optimization
Keywords:
flow matching, generative models on Lie groups, validity-aware generation

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_ReX8OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper asks whether properties of the paths used to train flow matching also hold for the trajectories the model actually generates. It shows that exact population matching preserves the distribution at each time and expected integrated state costs, while trajectory properties such as hitting probabilities and endpoint pairing can change. It also studies when integrated costs can bound endpoint violations, with experiments on tori, rotations, and molecular reconstruction.

Overall, I feel this is a careful analysis paper with a useful message, but the contribution feels somewhat incremental.

Critical strengths

The distinction between conditional paths, the population flow, and the fitted generator is useful. The circle example is particularly clear: it shows that trajectory statistics can change even with perfect regression and exact integration. I appreciated the separate pairing and path-design controls, the ten-seed evaluations, and the confirmatory entry-probability study with new seeds and multiple-testing correction. The paper also avoids treating lower path cost as automatic evidence of better endpoint fidelity.

Critical weaknesses

I’m not fully convinced by how much new insight is added beyond existing flow-matching theory. Rectified Flow already explicitly discusses preserving time marginals while changing the joint distribution of trajectories and the endpoint coupling. The circle counterexample and validity analysis are useful additions, but I would like a clearer explanation of what they teach us beyond that established distinction.

The learned-model results are harder to interpret than the analytic example. As the authors acknowledge, differences between training and generated trajectories can also come from fitting error. However, the capacity study in paper only tracks integrated cost, rather than the path-event differences. This leaves me unsure how much of the observed difference in trained models is actually caused by marginalization, which is central to the paper’s motivation.

Questions for authors

In Section 3.1, what is the main new conclusion about validity beyond the known distinction between matching time marginals and matching full trajectories? The circle example makes this distinction clear, but I would like to understand which existing assumptions or claims about flow matching it changes.

In Section 4.1, GAOC-FM (OT) reduces the accumulated state penalty along generated paths compared with OT-RFM. Does this mean fewer paths enter high-cost regions, or that they spend less time there or encounter smaller penalties? Could you clarify whether the lower cost also translates into fewer generated trajectories violating the validity criterion?

4 comments
AuthorsOpenReview ↗

Response to Reviewer ReX8 (1/2): attribution

Thank you. Our response comes in two parts because of the comment length limit. This part covers attribution and the meaning of lower cost; part 2 below covers novelty relative to Rectified Flow.

Fitting error or marginalization. You note that differences between training and generated paths could come from fitting error, and that the capacity study tracks only integrated cost, so it is unclear how much of the difference is due to marginalization.

You are right that the capacity plot in the PDF cannot attribute the hitting gap. For this response we measured the population field separately from the fitted student on the primary T2\mathbb{T}^2 tasks.

The oracle estimates density and velocity flux from 1.6×1071.6 \times 10^7 teacher paths per seed, with the training OT block size of 256, and integrates their ratio without fitting a student network. OT-RFM has no teacher network either; for GAOC-FM (OT) the learned teacher is frozen.

In the infinite-sample limit, Gaussian smoothing with bandwidth hh gives the population field of the teacher family with a static offset. The implemented oracle adds sampling, grid and integration error, so we compare it with the equally offset teacher and apply the declared marginal, fidelity and stability checks.

Write J=teacherh−teacherJ = \mathrm{teacher}_h - \mathrm{teacher}, Δpop=oracleh−teacherh\Delta_{\mathrm{pop}} = \mathrm{oracle}_h - \mathrm{teacher}_h and Δres=student−oracleh\Delta_{\mathrm{res}} = \mathrm{student} - \mathrm{oracle}_h, so that the observed gap is exactly J+Δpop+ΔresJ + \Delta_{\mathrm{pop}} + \Delta_{\mathrm{res}}. Here Δres\Delta_{\mathrm{res}} contains finite fitting together with any remaining oracle and numerical error.

Table R3. Seeds 10–19, 16,384 evaluation paths per seed; entry conditional on starting below τ=1\tau = 1; marginal 95% seed-bootstrap intervals.

Task and methodTeacherOracleh\mathrm{Oracle}_hStudentΔpop\Delta_{\mathrm{pop}} [95% CI]Δres\Delta_{\mathrm{res}} [95% CI]
Wide, OT-RFM0.03090.00000.0047−0.0311-0.0311 [−0.0324,−0.0296][-0.0324, -0.0296]+0.0046+0.0046 [+0.0031,+0.0064][+0.0031, +0.0064]
Wide, GAOC-FM (OT)0.02880.00000.0045−0.0291-0.0291 [−0.0304,−0.0278][-0.0304, -0.0278]+0.0045+0.0045 [+0.0032,+0.0057][+0.0032, +0.0057]
Narrow, OT-RFM0.02470.00040.0082−0.0248-0.0248 [−0.0255,−0.0240][-0.0255, -0.0240]+0.0078+0.0078 [+0.0058,+0.0099][+0.0058, +0.0099]
Narrow, GAOC-FM (OT)0.01140.00050.0044−0.0116-0.0116 [−0.0120,−0.0111][-0.0120, -0.0111]+0.0039+0.0039 [+0.0031,+0.0047][+0.0031, +0.0047]

The teacher column is unshifted; the offset used in Δpop\Delta_{\mathrm{pop}} changes mean entry by at most 0.0006. All eight Δpop\Delta_{\mathrm{pop}} tests (τ=0.5\tau = 0.5 and 11) have pHolm=0.031p_{\mathrm{Holm}} = 0.031 in the declared family of 16 tests, and at τ=1\tau = 1 the student lies between the oracle and the teacher. On the primary T2\mathbb{T}^2 tasks, marginalization is therefore the dominant source of the reduction, and the student residual works against it.

Every primary T2\mathbb{T}^2 cell passes the declared checks on integrated cost, endpoint W2W_2, and stability in sample size and bandwidth, and the basins null check is zero. In an exploratory check against the numerically integrated closed-form T2\mathbb{T}^2 RFM field, the kernel estimate of Δpop\Delta_{\mathrm{pop}} differs by at most 0.0022 in the seed mean and 0.0054 for a single seed. These are finite-resolution checks.

The exploratory T8\mathbb{T}^8 Monte Carlo field raises low effective-sample-size flags and fails the originally tabulated fidelity check. A later matched-sample-size comparison is reassuring, but the flags stand, and we do not use T8\mathbb{T}^8 for attribution. The limitations section will say that attribution is established for the validated T2\mathbb{T}^2 tasks.

What lower integrated cost means. You ask whether the lower state penalty of GAOC-FM (OT) relative to OT-RFM in §4.1 comes from fewer entries into high-cost regions, shorter stays or smaller penalties, and whether it means fewer trajectories violate the validity criterion.

On the T2\mathbb{T}^2 ridges, occupation of the high-cost region falls sharply while terminal violation does not, and the change comes from paths that start inside the region. The existing ten-seed E1 panel (descriptive) separates the two. At τ=0.25\tau = 0.25 on the wide ridge, moving from OT-RFM to GAOC-FM (OT) lowers mean occupation from 0.2409 to 0.0684, while terminal violation is 0.0409 against 0.0397. On the narrow ridge, occupation goes from 0.0532 to 0.0136 and terminal violation is 0.0027 for both. In the exploratory transfer records, no path enters the τ=0.25\tau = 0.25 region on either ridge at the scoring resolution.

On T8\mathbb{T}^8, the joint probability of starting below τ=1\tau = 1 and later entering falls from 0.221 to 0.133 (unadjusted p=0.002p = 0.002). This is a joint probability, unlike the conditional entry in Table R3. Integrated cost alone cannot separate these mechanisms.

AuthorsOpenReview ↗

Response to Reviewer ReX8 (2/2): novelty relative to Rectified Flow

Novelty relative to Rectified Flow. You note that Rectified Flow already separates preserving time marginals from changing the trajectory law and the endpoint coupling, and ask what the circle example and the validity analysis of §3.1 add, and which assumption or claim about flow matching they change.

We agree that marginal preservation and the change of endpoint coupling are established in Rectified Flow. The revision also credits the length and kinetic-energy inequalities to the rectification and interpolant literature and to conditional Jensen. What we add is the transfer distinction below and the controlled attribution experiment in part 1 of this response.

The hitting distinction can be stated with a fixed hazard. The construction is new and analytical, separate from the trained experiments that follow it.

Proposition. There are two endpoint-pinned shortest-geodesic teacher families on S1S^1 with the same one-time marginals, expected length, kinetic energy and teacher hitting rate for the same hazard, whose population flows have different hitting rates. For continuous teacher and population paths with equal one-time marginals, almost-sure avoidance of any open set transfers in both directions.

Construction and proof. Set T=1T = 1, let X0X_0 be uniform on S1S^1, and fix A=(0,1.3π)A = (0, 1.3\pi). Define γt=X0+Dt mod 2π\gamma_t = X_0 + Dt \bmod 2\pi, with DD independent of X0X_0. In family F+F_+, DD is 00 or 0.9π0.9\pi with equal probability; in family F0F_0, DD is 00 with probability 1/21/2 and ±0.9π\pm 0.9\pi with probability 1/41/4 each. The endpoint pair determines DD, and every nonconstant arc is a shortest geodesic.

In both families γt\gamma_t is uniform for every tt, E∣D∣=0.45π\mathbb{E}\lvert D\rvert = 0.45\pi and ED2=0.405π2\mathbb{E} D^2 = 0.405\pi^2. The complement of AA has width 0.7π0.7\pi, so every moving teacher that starts outside enters AA, and every stationary one stays outside: both conditional teacher entry rates are 1/21/2. Given any current state, DD keeps its original law, so the population angular velocities are ED=0.45π\mathbb{E} D = 0.45\pi and 00, with conditional entry rates 0.45/0.7=9/140.45/0.7 = 9/14 and 00. Counting starts inside AA, the hitting probabilities are

Pteacher=0.825 (both families),Pflow(F+)=0.875,Pflow(F0)=0.650.P_{\mathrm{teacher}} = 0.825 \text{ (both families)}, \qquad P_{\mathrm{flow}}(F_+) = 0.875, \qquad P_{\mathrm{flow}}(F_0) = 0.650.

Hitting can thus move in either direction with the hazard and these teacher summaries fixed; what differs is the endpoint coupling.

For avoidance, a continuous path that enters an open set spends positive time in it. Equal one-time marginals give equal expected occupation by Tonelli, so one law enters with probability zero if and only if the other does. This concerns continuous population paths, not fitted fields, sampled nodes or contact with a closed boundary.

Trained experiments. On the area-uniform annulus 1≤∣x∣≤21 \le \lvert x\rvert \le 2, the three arc families we trained have known rotational population fields. Conditional entry rates over ten seeds with 16,384 starts per seed (the exact column is analytical):

Conditional turnsHazard widthTeacher estimateExact population entryStudent estimate
00 or +1.5π+1.5\pi, equal weightsπ\pi0.5020.7500.744
00 or +0.9π+0.9\pi, equal weights1.3π1.3\pi0.5029/149/140.641
±0.75π\pm 0.75\pi, equal weightsπ/2\pi/20.50100.001

All three primary student–teacher contrasts have pHolm=0.006p_{\mathrm{Holm}} = 0.006, so trained generators follow the predicted direction in both cases. The hazard differs across rows, and F0F_0 itself has not been trained. The first family turns by 1.5π1.5\pi, so its arcs are not shortest geodesics.

The registered avoidance diagnostic uses the smaller open disc ∣x∣<0.9\lvert x\rvert < 0.9, which neither teachers nor the population flow enter; the largest student rate into that disc over seeds is 0.23%.

Which inference changes? Teacher hitting rate, marginals and length together do not determine the generator's hitting rate, whereas zero entry into an open set is guaranteed at the population level. Hitting therefore has to be measured on generated trajectories. Claims about marginals in earlier path-design work are unaffected, and the avoidance argument itself is elementary.

Do this distinction and the T2\mathbb{T}^2 attribution in part 1 resolve your concerns about novelty and attribution? If so, we would be grateful if you would reconsider your assessment; if not, we would like to know which objection remains.

Thank you for the response and I feel most of my concerns are addressed. Though I still think theoretical contribution is rather modest, the added construction and attribution experiments makes the paper more clear and convincing. I saw authors haven't incorporate the changes in revision, please add them and I'll raise my rating to 3 after that.

AuthorsOpenReview ↗

Response to Reviewer ReX8 about changes in the revised PDF

Thank you for the follow-up. As you asked, the construction and the attribution experiment are now in the revised PDF.

  1. The fixed hazard construction and avoidance of open sets are Proposition 1(b) and (c) in §3.1 (p. 4), proved in Appendix C.2 (p. 18). Table 1 (p. 5) is split to match, and §3.1 credits rectified flow and stochastic interpolants for marginal preservation and the change of coupling.
  2. The attribution experiment (our Table R3) is §4.2 and Table 3 (p. 8). Appendix H (p. 31) has all 16 tests, the sensitivity to sample size and bandwidth, the validity checks and the T8\mathbb{T}^8 caveat.
  3. Your question on what lower cost means is answered in Appendix F. Table 12 (p. 26) compares occupation with terminal violation, and Table 13 (p. 27) reports joint entry.

Two further points. Table 3 gives wide ridge teacher entry 0.0288, while the confirmatory Table 17 (previously Table 9) gives 0.024 for the same seeds. Table 17 uses the 1,024 stored evaluation pairs per seed, and the oracle protocol adds 15,360 fresh pairs from the same coupling. On the stored pairs alone, the oracle runs reproduce Table 17 exactly (teacher 0.024, student 0.005). Student entry is below teacher entry on both sets, by 0.019 and 0.024. The Table 3 caption and Appendix H state this.

We also trained the second family of Proposition 1(b) at the same hazard (Appendix I, Table 21, p. 33). Both teachers enter at 0.502, and the students enter at 0.641 and 0.003, against the predicted 9/14 and 0 (paired p=0.002p = 0.002, ten seeds). Our earlier caveat that F0F_0 was untrained no longer applies.

Thank you again for the careful review.

Reviewer_M7PcOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The authors observe that many validity properties in Flow Matching are defined for conditional paths, while the trained model learns a marginal vector field. This raises an important question: which properties of conditional paths are preserved after training? First, they introduce GAOC-FM, which fixes the source and target endpoints while allowing the paths between them to be modified to take detours. GAOC-FM (OT) further optimizes the endpoint coupling using Optimal Transport. The theoretical analysis shows that the one-time marginals can be preserved, while path properties such as hitting events and endpoint coupling are not necessarily preserved. The experiments first show that reducing path cost does not necessarily improve endpoint distribution quality. They then demonstrate that even with a speed bound, the resulting endpoint validity bound can be too loose to be useful. Finally, the experiments show that geometric constraints can limit possible detours and that uncalibrated validity evaluation can be misleading.

Critical strengths

The paper identifies the gap between conditional paths and marginal flows in Flow Matching. The experiments provide evidence of these differences. This offers a useful perspective for understanding the properties of Flow Matching.

Critical weaknesses

w1. The paper provides a useful analysis of which properties of conditional paths are preserved after marginalisation. However, the practical implications of the properties that are not preserved remain insufficiently explored. For example, it is unclear whether better preserving properties such as hitting events, endpoint coupling or other properties would improve the performance of the generated flow.

w2. Although the paper investigates speed caps and endpoint validity bounds, these interventions have limited effectiveness in the reported experiments. It would be helpful to include simple heuristic approaches or additional ablation studies to explore whether these limitations can be mitigated and how doing so affects generation quality.

Questions for authors
  1. The paper presents several theoretical results and experimental analyses, but the connections between them are not always clear. In particular, the transition from the theoretical analysis in Section 3 to the experiments in Section 4 feels abrupt. Could the authors improve the connection between these sections and explain more clearly how the experiments relate to the theoretical results?

  2. Can this analysis be extended to flow variants such as Flow Map and Mean Flow?

3 comments
AuthorsOpenReview ↗

Response to Reviewer M7Pc

Thank you. Your questions separate three objects the paper should keep apart: conditional paths, generated trajectories and the interpolation between sampler nodes.

Flow maps and MeanFlow. You ask whether the analysis extends to flow maps and MeanFlow.

An exact flow map of the population ODE reproduces the ODE's joint law on any time grid, but not what happens between nodes: straight interpolation between its outputs can lose the ODE's avoidance property.

We tested this on the annulus family that stays still or turns +1.5π+1.5\pi with equal probability. Its population ODE rotates by 0.75π0.75\pi, and neither the teachers nor the ODE enter ∣x∣<1\lvert x\rvert < 1. For a kk-step exact map with straight segments, a chord at radius rr reaches radius rcos⁡(0.375π/k)r\cos(0.375\pi/k). Since r2r^2 is uniform on [1,4][1, 4], the exact hole-entry probability is min⁡(1,[sec⁡2(0.375π/k)−1]/3)\min\bigl(1, [\sec^2(0.375\pi/k) - 1]/3\bigr). MeanFlow was declared in advance as a descriptive study (ten seeds).

StepsExact map with straight segmentsTrained MeanFlow, seed mean [range]
11.0000.990 [0.961, 1.000]
20.1490.199 [0.166, 0.241]
160.00180.040 [0.011, 0.093]

Across these step counts, MeanFlow's conditional hazard entry is 0.741–0.754, close to the population value 0.75 and above the teacher's 0.50. So on this task MeanFlow follows the population flow, and the hole crossings come from straight interpolation. We make no claim about sample quality or about flow maps in general. With one step, the discrete integral is a trapezoidal endpoint cost, to which Proposition C.4 still applies under its stated assumptions.

Would better preservation improve generation? You ask whether better preserving hitting events, endpoint coupling or other path properties would improve the generated flow.

Not the endpoint law: at the population optimum it is already the prescribed target for every admissible path family, and the revision states this right after eq. (9). In the trained models, lower cost does not come with a measurable gain in fidelity either; in §4.1 the paired W2W_2-difference intervals include zero on all four tasks.

Path properties matter in two places. When intermediate states are part of the output, teacher-level measurements overstate the generator: on the new R2\mathbb{R}^2 tasks, teacher length falls by 45–47% and generated length by 7–17% (Table R2, our reply to Reviewer cDML). In few-step sampling, the existing wide-ridge study gives a five-step Lie-Euler error of 0.081 for OT-RFM and 0.203 for RFM against each model's own converged rollout, which is numerical error, not endpoint distribution error.

The speed cap and the validity bound. You ask whether heuristics or ablations could mitigate their limited effect, and at what cost in quality.

We agree that the cap does not make the certificate informative at τ=0.25\tau = 0.25. Existing sweep: wide ridge, GAOC-FM (OT), five seeds, 200 steps; thresholds are ranges over seeds.

Cap BBFirst informative thresholdTerminal violation, τ=0.25\tau = 0.25MMD2×104\mathrm{MMD}^2 \times 10^4
11.290–1.3200.21964.1
21.593–1.6130.0813.8
31.861–1.8960.0372.6
4None in the tested grid0.0313.2

Tightening the cap from B=3B = 3 to B=1B = 1 lowers the threshold only from about 1.9 to 1.3, while terminal violation rises from 0.037 to 0.219 and MMD2×104\mathrm{MMD}^2 \times 10^4 from 2.6 to 64.1.

Inference-time correctors trade fidelity for validity. On the wide ridge, our adapted OC-Flow and PCFM controls reach terminal violation 0–0.002, with MMD2×104\mathrm{MMD}^2 \times 10^4 of 6.1–21.6 and 30.4–35.6 respectively (five seeds, descriptive; not a ranking of the released methods).

With stored endpoints, the direct DKW band gives 0.17 where the integral band gives 1 (Proposition C.4, equal sample size and confidence). This improves the certificate, not the generator or its cap.

From theory to experiments. You ask which experiment tests which theoretical result.

Table R1. In the revision this map organizes §4.

StatementEvidenceInterpretation
Eq. (9), population endpoint law§4.1 and Table 2Finite-model W2W_2 improvements are unresolved.
Prop. 1 and the hitting distinctionTables 8–9, new T2\mathbb{T}^2 oracle and annulusPrimary T2\mathbb{T}^2 gaps are mainly marginalization; either sign is possible.
Theorem 2 and Cor. 3§4.2, Fig. 5, Tables 5–6Valid but vacuous at the operating threshold.
Prop. C.4§4.2 and Fig. 5Direct endpoint estimation gives a tighter certificate.
Eqs. (3), (6), radius constraint§4.3 and Fig. 7At R=0.6R = 0.6, 32.2∘32.2^\circ is available where 42.3∘42.3^\circ is needed.

Does this separation of population transfer, numerical interpolation and endpoint certification address your concern about practical impact?

Thank you for the detailed response and additional experiments. The authors have addressed most of my concerns. However, I still have some concerns about the practical implications of this work for future Flow Matching research. Although the paper provides a useful understanding of which path properties are preserved after marginalization, it is still unclear to me how these findings can guide the development of better Flow Matching methods, beyond emphasizing the need to evaluate generated trajectories directly. I would appreciate a clearer discussion of the practical implications of these theoretical findings.

AuthorsOpenReview ↗

Response to Reviewer M7Pc about changes in the revised PDF

Thank you for following up.

it is still unclear to me how these findings can guide the development of better Flow Matching methods, beyond emphasizing the need to evaluate generated trajectories directly.

We agree that our first answer mostly said what to measure after training. The revision adds §4.5 and Table 4, which turn the transfer results into choices made before training, and Appendix J tests them on trained generators. The task moves a Gaussian to two modes past an open disc, with 24 designs and ten seeds each, scored by how often the generated trajectories enter the disc. No training loss contains entry, though the disc informs the path designs.

1. To avoid a region, keep it out of the support of every conditional path, then control leakage. Zero entry into an open set carries over to the population flow (Proposition 1(c)). A positive teacher rate does not have to. Generated entry and endpoint W2W_2 at 200 Heun steps (reference W2W_2 0.065) are below.

Conditional pathsEntryW2W_2
Straight, independent pairs0.9200.064
Straight, minibatch OT0.7810.062
Soft penalty, λ=6\lambda = 60.4650.064
Clearance 0.1 from the disc0.0170.182
Clearance with minibatch OT0.00060.122
Strong penalty, λ=30\lambda = 300.00010.070

Clearance enters 0.447 less often than the soft penalty (declared test, pH=0.012p_H = 0.012). Its teachers enter 0.01% of the time and its students 1.7%, so most of the gap is leakage from the finite model. The teachers only approximate exact avoidance, so here Proposition 1(c) is a design principle, not a certificate. Adding OT to the same paths lowered entry in all ten paired seeds and W2W_2 in eight (exploratory, unadjusted p=0.002p = 0.002 and 0.0120.012). With six nuisance dimensions, the declared addendum test gives 0.794 lower entry than straight OT paths (pH=0.006p_H = 0.006). The strong penalty beats clearance with OT on both entry and W2W_2, at higher kinetic energy (19.9 against 17.0). The lever is the support of the conditional paths, however it is obtained.

2. Coupling sets hitting rates, and timing sets occupancy at a given time. We trained the second family of the fixed hazard construction (Proposition 1(b)) under an addendum registered before the runs. Both families share the hazard, marginals, expected length and energy, and teacher entry (0.502). Their students enter at 0.641 and 0.003, close to the predicted 9/14 and 0 (paired p=0.002p = 0.002, Appendix I). In the design study, OT pairing alone cuts the entry of straight paths from 0.920 to 0.781 (pH=0.012p_H = 0.012). Keeping the routes and the teachers' hitting rate (0.843) fixed and changing only when the routes pass a gate moved generated occupancy at t=1/2t = 1/2 from 0.632 to 0.316 (pH=0.012p_H = 0.012), as Eq. (9) predicts.

3. Two declared predictions failed. We expected the entry rates of soft penalty paths with a random side not to carry over. They did. Students stayed within about 0.003 of their teachers on average (p=0.57p = 0.57 and 0.280.28), so this task gives no case of a positive rate failing to transfer. Those cases come from Proposition 1(b) and the oracle in §4.2. The rule is unchanged. A rate may carry over, but nothing guarantees it, so measure it on generated paths.

4. With few sampler steps, choose the step count and interpolation together with fidelity. With one Euler step, clearance with OT enters only 0.029 of the time but has W2W_2 1.26, so choosing by entry alone picks a poor sampler. No planar design met both declared bounds, on entry and on W2W_2, at the Euler step counts we tested (Table 22).

Correction on MeanFlow. Our first reply said the hole crossings come from straight interpolation. That is true only with one or two steps. At 16 steps, 0.039 of the 0.040 entry (97%) comes from paths with a learned node inside the hole, so these are errors of the learned map. Circular arcs between exact nodes remove chord crossings when the geometry is known, but cannot fix a learned node inside the hole (Appendix L, Table 26). We apologize for the error.

For terminal validity the advice is unchanged. Compare inference corrections by their fidelity cost (Table 10) and certify with the direct endpoint band (§4.3).

On scope, this is one planar task with deliberately simple designs, meant to test the levers. We claim no state of the art, and GAOC-FM still has no established fidelity advantage. The levers apply most directly where flow time is execution time, as in Streaming Flow Policy, which we have not run.

We are happy to clarify any of these points further.

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

This paper investigates which properties of conditional trajectories in flow matching are preserved by the learned vector field, particularly for trajectory validity on Lie groups, motivated by the fact that flow matching learns a marginal vector field by averaging conditional velocities, which may not preserve the properties of individual trajectories. To investigate this, the paper proposes GAOC-FM, which learns corrections to geodesic paths by minimizing a cost objective, and then trains a flow matching model on these trajectories. The authors theoretically show that population flow matching preserves marginal distributions and expected integrated state costs, but does not necessarily preserve path dependent properties. They also show that low cost alone does not guarantee terminal validity, and derive bounds under additional conditions on trajectory speed or energy. They evaluate the method on synthetic torus and SO(3) tasks against, and conduct additional experiments on molecular reconstruction. On the torus tasks, GAOC-FM reduces integrated cost compared to OT-RFM but does not improve final distribution generation quality. Furthermore, the teacher and generated trajectories show differences in their probabilities of entering high cost regions. However, the proposed terminal validity bounds and molecular experiments mainly demonstrate negative results, such as potential artifacts in measuring physical validity.

Critical strengths
  1. The paper identifies an important issue in flow matching, which is that properties of conditional trajectories used during training may not be preserved by the learned vector field.
  2. The paper provides theoretical results that characterize which properties of conditional trajectories are preserved under flow matching. In particular, the authors show that population flow matching preserves marginal distributions and expected integrated state costs, but may not preserve path dependent properties such as hitting events and path length. They also derive bounds on terminal validity under assumptions on the regularity of generated trajectories.
Critical weaknesses
  1. The paper is poorly organized and difficult to follow, especially in the experimental section. While the paper claims that the overall aim is to investigate which properties of trajectories transfer to the generated flow, the experiments largely address questions that are not clearly connected to this goal. For example, Section 4.4 studies molecular reconstruction artifacts, which has little to do with trajectory preservation. Additionally, Sections 4.3 and 4.4 provide very little explanation of their experimental setups in the main text, making it difficult to understand what the experiments are actually evaluating, and what even is the underlying benchmark. These sections also do not adequately demonstrate what they are investigating. For example, section 4.2 appears evaluate the practical usefulness of theorem 2 and section 4.3 appears to be ablating the hyperparameter RR in the geodesic correction bound, but the notation, the task, and the metric are not clearly explained. For example, where is the context for "At the central rotation crossing" and "Under the declared overlap rule", which are the opening sentences for these sections? Overall, the papers needs major revision to better explain the experimental setups and clearly indicate the theoretical contributions, clarify what each experiment aims to demonstrate, and establish a clearer connection between the results and the paper's central research question. Furthermore, the paper also does not clearly articulate or demonstrate the benefits of preserving trajectory dependent properties in generation paths.
  2. The bigger weakness is that the paper does not establish a sufficiently significant theoretical or empirical contribution for an ICLR paper. While the authors explicitly state that their main contribution is analyzing trajectory behavior instead of proposing a new path optimization method, the experiments are still mostly setup to compare GAOC with other baselines and neither the theoretical results nor the experiments demonstrate sufficiently novel or useful insights. In particular
  • The main theoretical result is that conditional flow matching preserves marginal distributions and expected integrated costs, but not necessarily individual trajectory properties. However, this is a relatively expected consequence of averaging conditional velocities, and the paper does not propose a way to address this limitation. In Section 4.1, the results show that optimizing conditional paths using a cost objective reduces the same cost in the generated trajectories, which is largely expected from the objective. Nevertheless, GSBM achieves lower integrated cost on three of the four torus tasks, so the experiments do not establish a clear advantage of GAOC-FM over existing approaches. As a result, the paper mainly identifies an intuitive issue for flow matching but does not propose a resolution to it.
  • Theorem 2 provides bounds on terminal validity under some regularity assumptions, but Section 4.2 shows that these bounds are uninformative, and applying the speed cap even increases integrated cost of the paths. The current experiments suggest that the theoretical results in Theorem 2 do not translate into practical benefits. This further undermines the contribution of the paper, especially since the derivation of Theorem 2 and its corollaries is relatively straightforward.
Questions for authors
  1. The problem for preserving trajectory level properties in flow matching is not limited to Lie groups. Are there any particular reasons the author do not evaluate similar experiments in the Euclidean case?
4 comments
AuthorsOpenReview ↗

Response to Reviewer cDML

Thank you for the detailed review. We agree that the method framing, the missing experimental context and the presentation of Theorem 2 obscure what the paper establishes; the general response lists the changes. Below we take up your remaining points in turn.

Predictability. You note that the main theoretical results follow predictably from averaging conditional velocities.

Marginal preservation and the one-sided length and energy inequalities are known, and the revision credits them. They give no one-sided law for hitting. Take a uniform initial angle on S1S^1 and the fixed arc A=(0,1.3π)A = (0, 1.3\pi). One teacher family stays still or turns +0.9π+0.9\pi, each with probability 1/21/2; the other stays still with probability 1/21/2 and turns ±0.9π\pm 0.9\pi with probability 1/41/4 each. Both use shortest circular geodesics and have uniform marginals at every time, expected length 0.45π0.45\pi and conditional entry 1/21/2 from outside AA. Their population flows enter AA with probability 9/149/14 and 00. The proof is in part 2 of our response to ReX8.

The construction is analytical. The annulus experiments test separately whether trained generators move in the predicted direction, and they do: student entry is 0.744 against a teacher rate of 0.502 in the increasing case, and 0.001 against 0.501 in the decreasing case. These two runs use hazard widths π\pi and π/2\pi/2, so they test direction and are not a matched-hazard comparison.

The second point is attribution. On the primary T2\mathbb{T}^2 tasks, a population oracle with no student regression enters at most 0.0005, against teacher rates of 0.011–0.031. Student residuals are positive, so fitting moves the students partly back toward the teacher. All declared T2\mathbb{T}^2 validity checks pass (Table R3, reply to Reviewer ReX8).

Benefit of preserving path properties. You note that the paper does not show what is gained by preserving them.

At the population optimum, every path family pinned to the same target has the same endpoint law, so preserving teacher path events cannot improve it. In practice, the generated trajectory has to be evaluated directly when intermediate states matter, and numerical error measured separately when few steps are used.

Table R2. Path-length reduction when independent pairing is replaced by minibatch OT (new R2\mathbb{R}^2 experiments, ten seeds, descriptive).

Task and path familyTeacher length reductionGenerated length reduction
Wide ridge, straight paths46.5%16.5%
Wide ridge, GAOC paths45.4%7.3%
Narrow ridge, GAOC paths46.9%15.4%

In each row the reduction measured on teachers is 2.8 to 6.2 times the generated one. For few-step sampling, the existing wide-ridge study gives a five-step Lie-Euler error of 0.081 for OT-RFM and 0.203 for RFM, each against the same model's converged rollout. This is solver error, not W2W_2 to the target.

Euclidean space. You ask why we did not test in Euclidean space, since the problem is not specific to Lie groups.

We added R2\mathbb{R}^2 ridge tasks with the same pipeline and four methods. Student entry falls in all 14 non-degenerate cells; of the 16 primary tests, 12 have pHolm=0.031p_{\mathrm{Holm}} = 0.031, two have 0.062 and two are null cells. The paired R\mathbb{R} experiment gives teacher entry 1.000 at reported precision and student rates of 0.023 and 0.018. Both are controlled, low-dimensional settings.

GSBM and Theorem 2. You note that GSBM has lower integrated cost on three of the four tasks, and that Theorem 2 is uninformative in §4.2.

GSBM-ZN does have lower mean cost on three tasks, and it has higher mean endpoint W2W_2 than GAOC-FM (OT) on all four. The comparison also changes capacity and amortization, so it supports neither a mechanism nor a general ranking.

We agree that Theorem 2 is vacuous at τ=0.25\tau = 0.25. Proposition C.2 shows that no terminal-risk bound depending only on expected cost can vanish uniformly as that cost goes to zero. Proposition C.3 shows that Theorem 2 is sharp over cost processes with a given mean and backward modulus, though not for a fixed target distribution. When endpoints are available, Proposition C.4 gives a better route: the DKW band is 0.17 where the integral band is 1. The revision presents this as a limitation and as guidance for evaluation.

Do the fixed-hazard construction and the T2\mathbb{T}^2 attribution experiment answer the concerns about predictability and fitting? If a specific mathematical or experimental issue lies behind the soundness score, we would be glad to address it directly.

AuthorsOpenReview ↗

Response to Reviewer cDML about changes in the revised PDF

As the discussion period closes, we note one update and where the revision addresses your points.

One statement in our reply is out of date. We said the trained annulus families had different hazards, so they tested direction but were not a matched comparison. We have now trained the matched partner (Proposition 1(b)), under an addendum registered before the runs. The two families share the hazard, marginals, expected length and energy, and teacher entry (0.502). Their students enter at 0.641 and 0.003, as predicted (9/14 and 0, exact paired p=0.002p = 0.002, ten seeds, Appendix I). Only the endpoint coupling differs, and it sets the hitting rate.

On your other points, the former §4.4 is now Appendix E, each experiment opens with its question and setup, and the Euclidean study is Appendix K. Theorem 2 is now preceded by the no-modulus and sharpness results (§3.2), and §4.3 reports that it is uninformative at our threshold, where the direct endpoint band gives 0.17. You wrote that the paper "does not propose a way to address this limitation". The new §4.5 derives design choices from the transfer results and tests them on trained models (see our reply to Reviewer M7Pc), including two predictions that failed.

I thank the authors for their rebuttal and revision. I believe the paper is clearer than before with the revision, and the new hazard and population oracle experiments provide stronger evidence that population marginalization itself can change hitting probabilities, instead of attributing these differences mainly to fitting errors. However, my main concern remains whether these findings provide sufficiently novel or useful insights for generative modeling. The new Section 4.5 looks into design choices such as obstacle clearance paths, OT coupling, and stronger cost penalties, but these are existing techniques, and it is unclear what additional benefits the theoretical analysis provides. For example, adding OT to clearance paths reduces generated obstacle entry to 0.06%, but still gives worse endpoint fidelity than straight paths with OT (W2 of 0.122 vs 0.062), while stronger penalties further reduce entry at the cost of higher kinetic energy. Two of the study's predictions also fail, and none of the tested few-step Euler configurations meets both the entry and fidelity targets. Furthermore, the terminal validity bounds from Theorem 2 remain uninformative. Overall, while I appreciate the improvements in the presentation and experimental validation, I still have concerns about whether the theoretical results lead to sufficiently meaningful contributions beyond characterizing a known limitation of conditional flow matching. I will update my assessment to reflect the writing changes given my remaining concerns.

AuthorsOpenReview ↗

Response to Reviewer cDML on the remaining concern about significance

Thank you for reading the revision and for updating your assessment.

these are existing techniques, and it is unclear what additional benefits the theoretical analysis provides

We agree that clearance paths, OT coupling and cost penalties are existing techniques, and we do not claim a new method. The analysis shows that before training, which choices preserve which properties in the generator. Conditional paths that keep the obstacle out of their support carry zero entry to the population flow (Proposition 1(c)), so only fitting and sampling leakage remain to control. Lowering a positive teacher rate gives no such guarantee (Proposition 1(b)), and that rate has to be measured on generated paths. The matched pair shows why teacher statistics cannot replace this measurement. The two families agree in hazard, marginals, expected length and energy, and teacher entry, yet their students enter at 0.641 and 0.003, as predicted (Appendix I).

We'd like to give a clarification on the price of each design. With the reference sampler of 200 Heun steps, the strong penalty (λ=30\lambda = 30) meets both thresholds of our registered protocol, entry below 1% and W2W_2 within 0.02 of the reference 0.065. It reaches entry 0.01% and W2W_2 0.070, at kinetic energy 19.9 against 16.1 for straight paths with OT (Table 22). Its teachers rarely enter (0.0001), which is close to the case Proposition 1(c) describes. The joint failure you point out holds for Euler sampling with few steps, and we agree that it remains open.

We also agree that Theorem 2 is uninformative at our threshold, as §4.3 states.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General response: what transfers, new evidence, commitments.

We thank all the reviewers. We agree that the submission does not make its contribution clear. The revision reframes the paper around what a designed conditional path does and does not tell us about the generator, and separates the known marginal-preservation result from our own analysis and experiments.

The central distinction

Under the population matching assumptions of Appendix C.2, three things transfer from the conditional paths to the generator: one-time marginals, state-only terminal validity and expected additive state–time costs. The marginal/coupling picture is due to Rectified Flow and related work, and we credit it. Expected length and kinetic energy obey one-sided inequalities by conditional Jensen. Hitting probabilities need a different argument.

Our reply to Reviewer ReX8 constructs two teacher families with one fixed hazard, the same one-time marginals, the same expected length and the same teacher hitting rate, whose population flows enter the hazard with probability 9/149/14 and 00. The same reply proves that almost-sure avoidance of an open set does transfer for continuous population paths. A positive teacher hitting rate therefore does not determine the generator, while zero probability carries over. The construction is analytical and separate from the trained experiments.

Evidence addressing the reviews

ConcernAnswer in the repliesScope
Could the observed gaps be fitting error?The T2\mathbb{T}^2 population oracle enters at most 0.0005, against teacher rates of 0.011–0.031 at τ=1\tau = 1. Student residuals point toward the teacher.All declared validity checks pass in the primary T2\mathbb{T}^2 cells.
Does this extend beyond compact groups?On R2\mathbb{R}^2, student entry is lower in all 14 non-degenerate cells, and 12 of 16 tests survive Holm adjustment.Two tests do not survive; two cells are zero against zero.
Can hitting increase as well as decrease?Annulus students reproduce both directions; all three primary contrasts have pHolm=0.006p_{\mathrm{Holm}} = 0.006.The trained families use different hazard arcs, listed in the reply.
Does it matter for flow maps?With straight interpolation between nodes, a one-step exact map crosses an avoided hole with probability one; trained MeanFlow does so with probability 0.990.Descriptive result on the annulus.
Is Theorem 2 useful at the operating threshold?No. It gives 1 where the direct endpoint DKW band gives 0.17.A limitation of the integral certificate.

The T2\mathbb{T}^2 oracle and the Euclidean primary tests use ten seeds and exact sign-flip tests, Holm-adjusted within their declared families, and their designs were committed before the runs. MeanFlow was declared separately as a descriptive study without hypothesis tests. Analyses of existing data and the RFM oracle extensions are exploratory. The T8\mathbb{T}^8 oracle is flagged by its validity checks, and we do not use it for attribution.

Revision commitments

  1. Adopt the title "What Transfers from Conditional Paths to Flow-Matching Generators? A Controlled Study on Lie Groups". Present GAOC-FM as a controlled path family, remove method-ranking language and confine its name to the appendix.
  2. Credit Liu et al. (2023), Q. Liu (2022), Albergo and Vanden-Eijnden (2023), and Shaul et al. (2023) for the marginal and kinetic results. Split Table 1 into marginal quantities, length/energy inequalities, hitting, and open-set avoidance.
  3. State the no-modulus and sharpness results, with their precise scope, before Theorem 2, and recommend direct estimation of terminal violation when endpoints are observed.
  4. Add the theory-to-experiment map (Table R1, reply to Reviewer M7Pc). Open each experiment with its question, the statement it tests and its setup. Bring the confirmatory transfer evidence into §4.1 and move §4.4 to the appendix.
  5. Include the new results and the previously omitted numerical-error, corrector and cap tables, and state that population endpoint correctness does not imply fidelity gains in finite models.

These commitments apply to the revision phase, we will keep the current submitted PDF unchanged. One new result goes against GAOC-FM, and we report it: on R2\mathbb{R}^2, GAOC students have endpoint W2W_2 of 0.47–0.53, against 0.40–0.44 for straight-path students. The evidence supports a transfer analysis, not a better generator.

AuthorsOpenReview ↗

Summary of revisions

We thank all reviewers for the discussion. We have uploaded the revised PDF. It makes the changes promised in our general response, and the reviewers who asked for each one are in parentheses.

  1. Framing (cDML). New title, "What Transfers from Conditional Paths to Flow-Matching Generators? A Controlled Study on Lie Groups". §2 presents our paths as a controlled family, and ranking language is removed from the text and tables. One change from our plan. We kept the label GAOC-FM in the main text so that every number matches our replies here. It names one member of the path family and is never ranked.
  2. Relation to rectified flow (ReX8, M7Pc). §3.1 credits rectified flow and stochastic interpolants for marginal preservation and the change of coupling, and Q. Liu (2022) and Shaul et al. (2023) for the length and energy inequalities. Table 1 now separates marginals, length and energy, hitting, avoidance of open sets, and maxima and winding. Eq. (9) is followed by the statement that a better conditional path promises no fidelity gain for a finite model.
  3. Fixed hazard construction (ReX8, cDML). Proposition 1(b) and (c), proved in Appendix C.2. Both families are now trained (Appendix I, Table 21). At the same hazard and teacher entry 0.502, the students enter at 0.641 and 0.003, as predicted.
  4. Attribution on T2\mathbb{T}^2 (ReX8). New §4.2 and Table 3. Appendix H has all 16 tests and the sensitivity and validity checks.
  5. Theorem 2 (cDML, M7Pc). §3.2 states the no-modulus and sharpness results before Theorem 2 and recommends direct estimation of terminal violation when endpoints are observed.
  6. Structure of §4 (M7Pc, cDML). Table 2 maps each experiment to the statement it tests (our Table R1), and each subsection opens with its question and setup. §4.1 summarizes the confirmatory transfer evidence. The molecular study moved to Appendix E.
  7. Design implications (M7Pc, cDML). New §4.5 and Table 4, tested in a planar obstacle study (Appendix J), including two declared predictions that failed.
  8. Further results (cDML, M7Pc, ReX8). Euclidean controls (Appendix K), MeanFlow (Appendix L), occupation and joint entry (Tables 12 and 13), numerical error (Table 14) and corrector strength (Table 10).

One correction to our first reply to M7Pc. Straight interpolation explains MeanFlow's hole crossings only with one or two steps. At 16 steps, 97% of crossing paths have a learned node inside the hole (Appendix L, Table 26).

Numbers cited in our first replies have moved. The revision has Tables 6, 5 and 7 (previously 2, 3 and 4), Tables 8 and 9 (previously 5 and 6), Table 15 (previously 7), Tables 16 and 17 (previously 8 and 9), and Figures 6 and 5 (previously 5 and 6). §4.3 and §4.4 were §4.2 and §4.3, and Appendix E was §4.4. Our reply tables R1 and R3 are Table 2 and Tables 3 and 18, and Table R2 is in the text of Appendix K. Eq. (9), Theorem 2, Corollary 3 and Propositions C.2 to C.4 keep their numbers.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.