acceptodds
Under review as a conference paper at ICLR 2027

When Is a Conditional Path Already Optimal? A Criterion for Flow Matching on Lie Groups

Abstract

Under kinetic optimality, flow matching does not leave the conditional path free: that path is the geodesic of the metric the loss induces. On , where models now generate protein backbones, molecular poses and robot actions, that geodesic generally requires a boundary-value solve at a cost comparable to a training step, so implementations substitute one of three closed-form constructions. Yet how far the substitute sits from the curve it replaces has not been reported. We give that deviation in closed form for group interpolation: second order in the endpoint amplitude, with leading term . A deviation alone does not identify which constructions need changing, so we also prove a criterion for when it vanishes. A factorized construction solves the geodesic equation of a block metric exactly when the endpoint rotation axis is an eigenvector of . The condition is readable off the loss, and every body-frame-invariant loss with a block-diagonal Gram operator satisfies it by Schur's lemma. For a given axis, what decides the matter is alignment rather than anisotropy: rotating a metric's eigenbasis at fixed eigenvalues takes the construction from solving the geodesic equation exactly to departing from it. Reading six released loss configurations at source level, in the coordinates where each loss is computed, we find four that already satisfy the criterion, without having been built for it. Where it holds, the departure stays at the solver floor and does not grow with amplitude. Where it fails, the construction sits at second order from the geodesic, and each level of a closed-form correction removes one further order. Whether the departure is zero is therefore decidable before any training. On a robot action head the criterion accepts, training instead on the geodesic of a frozen approximation to its metric changes the error on training episodes by less than percent, although the two paths differ by percent of that error in translation and percent in rotation.

Primary area:
generative models
Keywords:
flow matching, Lie groups, SE(3), conditional path, Riemannian geometry, geodesics, protein backbone generation, robot policies

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_gh1bOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2(+1)
Presentation
2
Soundness
2(+1)
Read review
Paper summary

The paper asks when the closed-form conditional paths used in SE(3) flow matching are the geodesic of the metric that the training loss induces, and how far they are from it otherwise. It considers three paths: group interpolation, algebra interpolation, and the factorized SO(3) x R^3 path.

  • Thm 1: the leading gap of group interpolation is tau(1−tau)/2·Lambda(u).
  • Thm 2: group and algebra interpolation differ only at third order.
  • Prop 4: the factorized path is the geodesic of diag (J, beta) exactly when the rotation axis is an eigenvector of J.
  • Prop 7: closed-form corrections (Alg. 1), each removing one order.

of six released configurations, the paper finds that four satisfy the criterion. One violates it (FrameFlow above its auxiliary-loss gate), and one induces no quadratic metric. On SmolVLA/LIBERO, training on corrected paths or on the exact geodesic of a frozen metric changes in-sample regression error by less than 0.3%.

Critical strengths

The paper seems to be written by authors well versed in differential geometry.

Critical weaknesses

"Optimal" has no demonstrated consequence. Are the models better trained, are they trained more stable, do they converge faster? At no point is this clear to me while reading the paper. Maybe the authors could highlight an actionable consequence from their paper that either provides a new insight or leads to an improvement procedure for training machine learning models. Maybe there would be something to be gained form a new interpolation scheme for models operating on Lie groups and algebras?

Questions for authors

See the weaknesses.

7 comments
AuthorsOpenReview ↗

Response to Reviewer gh1b (1/2): the actionable consequences of optimality, and the interpolation scheme

We thank the reviewer for an accurate summary of Theorems 1 and 2, Propositions 4 and 7, Algorithm 1 and the audit of six released configurations. Whether a conditional path on SE(3)\mathrm{SE}(3) needs changing could previously be settled only by training two models and comparing them, and the paper turns this into a decision read from the loss before training (§6, lines 514 to 515). Training confirms the decision in both directions: where the criterion accepts the released path, a different path brings no detectable gain, and where it rejects it, a path closer to the geodesic trains to a lower held-out objective.

On the ratings. We respectfully note four points in the review itself.

(i) Soundness is rated 1, "significant concerns about the paper's soundness that do not appear easy to address", but the review names no derivation, proof or measurement that it considers incorrect, and its summary restates Theorems 1 and 2 and Propositions 4 and 7 without objection. We would be grateful to learn the specific concern, so that we can answer it.

(ii) Informativeness is rated 1, "I learned little or nothing new", while the summary lists the closed-form gap of Theorem 1, the alignment condition of Proposition 4 used as a criterion for released flow-matching paths, and the corrections of Proposition 7, none of which, to our knowledge, prior work on flow matching reports (Appendix A.1).

(iii) The weakness asks whether "a new interpolation scheme" could help, and the summary already lists one: "Prop 7: closed-form corrections (Alg. 1), each removing one order".

(iv) The strengths section comments on the authors' background but on none of the results, so the response below sets out what the results change.

W1. The actionable consequence, and new insight. Prior work shows that the choice of conditional path affects training in flow matching: optimal-transport paths train and sample faster than diffusion paths and generalise better (Lipman et al., ICLR 2023), simpler flows are more stable to train and faster at inference (Tong et al., TMLR 2024), and the curvature of the trajectories is directly related to the truncation error of the solver (Lee et al., ICML 2023). With the geodesic distance as premetric, Riemannian flow matching prescribes the geodesic as the conditional path (Chen and Lipman, ICLR 2024, Proposition 3.2), and on SE(3)\mathrm{SE}(3) released models substitute a closed-form construction, because the geodesic of the metric their loss induces needs a boundary-value solve. What the paper adds is when that substitution is harmless and what to do when it is not.

(i) For existing models, the criterion gives a verdict and a remedy. Four of six released configurations already train on the prescribed path, none by design (§5.1), so their developers need no boundary-value solve, which costs 21.9 to 158.1 ms per geodesic, from a fifth of a training step at 3.4 degrees to one and a half steps at 116 degrees, on forty cores (§5.3, lines 472 to 474). The one violation, FrameFlow above its auxiliary-loss gate, is traced to an uncentred backbone template that couples rotation and translation, and centring the template lowers the gap by a factor of 32 (§5.1).

(ii) For new models, the paper gives a design rule and a new insight. A loss invariant under a change of body frame, with a block-diagonal Gram operator, makes the factorized path the geodesic for every endpoint pair (Proposition 6). What decides optimality is the alignment of the metric's eigenbasis with the rotation axis, not anisotropy: at fixed eigenvalues, turning the eigenbasis by 45 degrees moves the same path from 1.8×10−131.8 \times 10^{-13} to 3.3×10−23.3 \times 10^{-2} Å from the geodesic (§5.2).

W1, continued. A new interpolation scheme. Algorithm 1 is such a scheme, and the criterion says where it is needed. Algorithm 1 and Eq. 11 correct a closed-form path in closed form, each level removing one order of the gap, at a fixed cost of a few Lie-algebra operations and one exponential (Appendix C.12). One level costs 38.0 µs per pair against 98.9 ms for a boundary-value solve, a factor of 2 601 at the median (Appendix D.6), and the fitted slopes of 1.000 for algebra interpolation and 2.023 and 3.001 for group interpolation corrected once and twice are orders two, three and four (Table 2). The criterion also selects the route: above FrameFlow's gate, correcting the released path in place lowers the gap by a factor of 2.8, while switching to group interpolation first raises it by a factor of 14.6 (§5.3).

AuthorsOpenReview ↗

Response to Reviewer gh1b (2/2): W1 continued, better trained, more stable, faster to converge

W1, continued. Better trained, more stable, faster? The criterion predicts the answer before training, and new experiments on held-out data confirm it in both directions. The new results are in the revised PDF, in Table 3 and Appendices D.12 and D.13, with all changes marked in blue.

(i) Where the criterion accepts the released path, a different path brings no detectable change in accuracy, stability or convergence. The review's summary reports the result of §5.4, a change of the in-sample error below 0.3 percent, which is the outcome the criterion predicts for a path it accepts. We repeated it out of sample. On the robot action head we retrained the two policies of Table 3, one on the released path and one on the exact geodesic of the frozen metric, two paths that differ by 3.7 and 5.9 percent of the policy's error, using 1 525 training episodes and seeds 42, 43 and 44, and evaluated each on the 168 held-out episodes. Pooled over seeds, the held-out errors differ by 0.021, −0.001-0.001, −0.001-0.001 and −0.035-0.035 percent in per-step translation, per-step rotation, composed translation and composed rotation, every 95 percent interval covering zero at a resolution of 0.057 to 0.164 percent. Their spread over seeds is the same, 0.03 to 0.26 and 0.03 to 0.23 percent of the mean, and so is their convergence: at every checkpoint from 5 000 to 25 000 steps the two means differ by less than one seed standard deviation, and both come within 1 percent of their final error at 20 000 steps. In closed loop on LIBERO-Spatial, with identical initial states, the success rates in percent with Wilson 95 percent intervals and the exact McNemar test are:

SeedReleased pathGeodesic of the frozen metricMcNemar pp
4230.6 [26.7, 34.8]29.8 [26.0, 34.0]0.74
4333.4 [29.4, 37.6]32.0 [28.1, 36.2]0.49
4432.4 [28.4, 36.6]30.4 [26.5, 34.6]0.34
Pooled32.1 [29.8, 34.5]30.7 [28.5, 33.1]0.20

These rates agree with the 35.5 percent of Appendix D.10, measured on 200 episodes (Wilson interval 29.2 to 42.3).

(ii) Where the criterion rejects the released path, a path closer to the geodesic trains to a lower held-out objective. On FrameFlow above its gate, the departure of the released velocity target is ρ=0.021\rho = 0.021 of the network's regression residual, against 5.6×10−155.6 \times 10^{-15} below the gate. We fine-tuned the released checkpoint for 2 300 steps with its training configuration unchanged, auxiliary weight one and the gate at t=1/4t = 1/4, on the released path and on a centred path that lies closer to the geodesic on all 1 000 training segments, with seeds 42, 43 and 44, and evaluated each model on 64 held-out SCOPe domains. Entries are the change of each model's held-out training objective, centred minus released, in percent, with 95 percent bootstrap intervals over domains.

Seedt<1/4t < 1/4 (criterion accepts)t>1/4t > 1/4 (criterion rejects)
420.01 [−0.24, 0.27]−1.65 [−2.46, −0.92]
430.07 [−0.16, 0.30]−0.82 [−1.60, −0.15]
440.22 [−0.17, 0.60]−2.08 [−2.96, −1.23]
Pooled0.10 [−0.06, 0.27]−1.52 [−2.16, −0.95]

Because each objective is the model's own, we also scored all checkpoints against common targets on identical held-out states. Above the gate the centred-path model is closer to the released target, on which only the released-path model was trained, by 1.76 percent [−2.41,−1.19][-2.41, -1.19] in the training objective, and to the velocity of the solved geodesic, by 0.51 percent [−0.77,−0.25][-0.77, -0.25]. Below the gate every difference covers zero.

Whether a change of conditional path can improve training is therefore decided from the loss before training, and the scale of the change is read from the released checkpoint before any alternative is trained: no detectable change where the criterion accepts the path, and a change of the order of ρ\rho where it rejects it. We hope that these answers, together with the interpolation scheme already in the paper and the new held-out experiments, address all of the reviewer's questions, and we would be grateful if the reviewer would consider raising the rating in light of them.

Thank you for the new experiments. The mathematics is correct, I stand corrected, and I raise my rating to 2. My concerns are significance, evidence, and presentation.

  1. The path does not matter. On SmolVLA, a path differing by 3.7–5.9% of the policy's error changes the trained policy by under 0.06%, with identical training curves (Table 17). Where the criterion fails, the path error is 2% of the network's own error (ρ = 0.021). The practical message is that existing paths need no change, which the title and framing do not convey.
  2. The effect is below ordinary training noise. On SmolVLA, the seed-to-seed spread exceeds the effect of the path (Table 16). On FrameFlow, the effect varies from 0.82% to 2.08% across seeds. The only violation exists because of the auxiliary weight and gate; the repository default has none (D.1). Learning rate, batch size or loss weights plausibly matter more, and no comparison is given.
  3. The positive result is confounded. Centring also changes the frame coordinates (l.375–377). The centred model gains more against the released target (1.76%) than against the geodesic (0.51%), the opposite of what the claimed mechanism predicts. Algorithm 1 is never trained, and no designability or scRMSD is reported.
  4. The harness is degraded. D.10 validates the harness by agreement with a published checkpoint scoring 38.5%, against about 90% reported for SmolVLA on LIBERO-Spatial.
  5. Presentation. Like NmqX, I find the paper and rebuttal read as machine-generated. Parsing them takes a lot of effort. Just in terms of writing I can not recommend this paper.
Reviewer_QUTVOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper quantifies how far the common closed-form conditional paths on Lie groups (group, algebra, and factorized interpolation) sit from the kinetic-energy-minimizing geodesic of a left-invariant metric. It gives the leading discrepancy in closed form (second order for group interpolation, third order between group and algebra interpolation), two levels of closed-form correction, and, for constant block metrics on SE(3)SE(3), an alignment criterion for when the factorized path is already geodesic. It then audits six released implementations and trains four policies on a robot action head.

Critical strengths

This paper makes the distinction between group, algebra, and factorized interpolation precise and useful. The asymptotic corrections are a strong contribution, offering explicit alternatives to repeated boundary-value solves. The policy experiment is carefully designed: it controls sampling noise and resamples episodes rather than chunks. In the isotropic SE(3)SE(3) case, the predicted second-, third-, and fourth-order error scaling agrees with my own calculation. This supports the formulas in that case, though it is not a verification of the general proofs.

Critical weaknesses

The identification of a regression loss with the kinetic metric used in this paper needs stronger justification.

There are several distinct objects here: a norm on velocity residuals, the pullback of that norm through an endpoint-prediction parameterization, and a Riemannian metric governing motion of the state. These need not coincide. For example, the public FrameFlow training code compares log⁡(Rt⊤R^1)/(1−t)\log(R_t^{\top}\hat R_1)/(1-t) with log⁡(Rt⊤R1)/(1−t)\log(R_t^{\top}R_1)/(1-t). Read as a velocity residual at the anchor RtR_t, where Jr(0)=IJ_r(0)=I, the induced metric is the identity. Read as the sensitivity to the predicted endpoint, it is Jr((1−τ)ω)−⊤Jr((1−τ)ω)−1J_r\big((1-\tau)\omega\big)^{-\top}J_r\big((1-\tau)\omega\big)^{-1}, which is the form the paper adopts. These are different objects, and the manuscript does not derive either as the metric on the evolving state. For FrameFlow the two readings happen to give the same verdict; for SmolVLA they do not. This is a concern about the practical interpretation rather than a counterexample to the fixed-metric theorems, but the audit rests on it. The manuscript should derive the connection explicitly from the training objective.

Relatedly, the premise and the policy experiment provide limited support for the main narrative. The claim that under kinetic optimality the conditional path must be the geodesic of the metric induced by the loss is not derived: Shaul et al. (2023) concern the marginal path, and Appendix C.1 explicitly disclaims any statement about the marginal. Since any conditional path yields a valid marginal, the manuscript should explain why the conditional action is the right quantity to minimize. The policy experiment does not settle this, as it is evaluated on training episodes (Appendix D.9) and compares against a frozen-metric geodesic while the released path is already optimal under the paper’s own pullback reading.

The summary of the implementation audit also requires careful calibration. Two of the four configurations reported as optimal have isotropic metrics, for which the criterion holds trivially; SmolVLA satisfies it by construction, since a straight line in the chart of its loss is the pullback geodesic; and the FrameFlow verdict below the gate relies on a frozen metric, which does not establish the geodesic equation of the varying one. The audit should separate exact fixed-metric results, fixed-chart pullback results, and frozen-metric diagnostics.

Questions for authors
  • For one audited implementation, could you derive the metric from the actual velocity-regression objective, tracking the endpoint parameterization and its Jacobian?
  • Which verdicts in Table 1 are exact without freezing? For FrameFlow below its gate, can the Euler--Lagrange residual of the original varying metric be evaluated?
  • Can the policy comparison be run on held-out episodes with repeated seeds, and can a configuration that violates the criterion (FrameFlow above its gate, coupled vs.\ centered template) be trained?
8 comments
AuthorsOpenReview ↗

Response to Reviewer QUTV (1/3): W1 and Q1, the metric derived from the training objective

We thank the reviewer for finding the distinction between group, algebra and factorized interpolation precise and useful, the corrections a strong contribution that replaces repeated boundary-value solves, and the policy experiment carefully designed, and for checking the second, third and fourth order scaling against an independent calculation. Beyond the isotropic case the reviewer checked, Theorem 2 is verified in exact arithmetic in the free Lie algebra (Appendix C.3), and we have since checked Theorem 1 and both correction levels on SO(4)\mathrm{SO}(4) with an anisotropic metric and on SL(2,R)\mathrm{SL}(2,\mathbb{R}), with the same orders as on SE(3)\mathrm{SE}(3). The derivation the reviewer asks for confirms every verdict of Table 1 under the metric the training objective induces, without freezing.

W1 and Q1. The metric derived from the training objective. We agree that a norm on velocity residuals, its pullback through an endpoint parameterization, and a metric on the evolving state are distinct objects, and that the submitted Appendix C.8 stated their link in one line. The link comes from Riemannian flow matching itself: its loss is the squared norm of the velocity residual at the current state in a metric gg (Chen and Lipman, ICLR 2024, Eq. 10), and with the geodesic premetric its conditional path is the geodesic of the same gg (their Proposition 3.2). Deriving gg from each audited objective gives the same verdict under both readings, and in each configuration the agreement has a specific reason.

(i) FrameFlow below its gate. The loss compares log⁡(Rt⊤R^1)/(1−t)\log(R_t^{\top} \hat R_1)/(1 - t) with log⁡(Rt⊤R1)/(1−t)\log(R_t^{\top} R_1)/(1 - t). Read as a velocity residual at RtR_t, the form is the identity, a bi-invariant metric under which every one-parameter subgroup is a geodesic (Appendix C.4). Read as sensitivity to the predicted endpoint, it is the form given in the review, and this form is itself a metric on the evolving state: the Euclidean pullback through the fixed logarithm chart Φ1(R)=log⁡(R⊤R1)\Phi_1(R) = \log(R^{\top} R_1) anchored at the data endpoint. Differentiating R⊤R1=exp⁡(Φ^1)R^{\top} R_1 = \exp(\hat\Phi_1) along R˙=Rξ^\dot R = R\hat\xi gives Jr(Φ1)Φ˙1=−exp⁡(−Φ^1)ξJ_r(\Phi_1)\dot\Phi_1 = -\exp(-\hat\Phi_1)\xi, and since Jr(Φ1)J_r(\Phi_1) commutes with exp⁡(Φ^1)\exp(\hat\Phi_1), ∥Φ˙1∥2=ξ⊤Jr(Φ1)−⊤Jr(Φ1)−1ξ\lVert \dot\Phi_1 \rVert^2 = \xi^{\top} J_r(\Phi_1)^{-\top} J_r(\Phi_1)^{-1} \xi. Along the released path R0exp⁡(τω^)R_0\exp(\tau\hat\omega) this chart reads Φ1=(1−τ)ω\Phi_1 = (1 - \tau)\omega, the argument in the reviewer's expression, so the path is a straight segment in a chart where the metric is Euclidean, and by Lemma 16 it is the exact geodesic of the unfrozen metric for ∥ω∥<π\lVert \omega \rVert < \pi. Corollary 5, which now states the anchor at either endpoint, covers this case (Appendix C.6, lines 1655 to 1656). We agree that the submitted Table 13 reached this verdict through the frozen form. The derivation above reaches it without freezing, and the entry of Table 13 now states the pullback. The Euler-Lagrange residual of this varying metric is computed directly under Q2 below.

(ii) SmolVLA. The network predicts the velocity in one fixed chart of normalised increments, in which the endpoint is affine in the velocity, so both readings give the same constant form, and the metric on the state is its flat pullback. The released straight path is its exact geodesic (Lemma 16, Appendix C.8 (i)). The differing verdict mentioned in the review comes from the left-invariant freeze, which is an approximation rather than a reading of the objective. Appendix C.8 states this in its last paragraph, "a fixed-chart loss is therefore judged under its pullback" (lines 1885 to 1886), and the paper uses the freeze only as the comparator of §5.4.

(iii) FrameFlow above its gate. The auxiliary loss compares predicted atom positions R^1ba+p^1\hat R_1 b_a + \hat p_1, so its form is defined on the endpoint and contains a constant coupling block, the hat map of ∑aba\sum_a b_a (Appendix D.1). The violation therefore does not depend on freezing, and Part 3 (ii) measures it: the released path departs from the geodesic by 0.089 to 0.091 Å under the metric of §5.3 and by the same amount under the unfrozen metric.

We have added the derivation of (i) to (iii) to Appendix C.8, which now lists FrameFlow below its gate under Corollary 5.

AuthorsOpenReview ↗

Response to Reviewer QUTV (2/3): W2 on the conditional action and the policy experiment, W3 and Q2 on the audit without freezing

W2. Why the conditional action. We agree that Shaul et al. study the marginal path, and that our opening sentence stated its link to the conditional path too briefly. The results themselves concern the conditional path (Appendix C.1), and none of them changes. The abstract and the introduction now state the premise on this basis. Among Gaussian conditional paths in Euclidean space, the conditional optimal-transport path (Cond-OT) minimises the conditional kinetic energy, which bounds the marginal one from above (Shaul et al., ICML 2023, §3.1 and §4). The bound holds on any Riemannian manifold, as the marginal velocity is the conditional mean of the conditional velocities and the squared norm is convex (Appendix C.1). In Riemannian flow matching with the geodesic distance as premetric, the conditional path is the geodesic of the metric (Chen and Lipman, ICLR 2024, Proposition 3.2), the curve of least conditional action between the two endpoints, and SE(3)\mathrm{SE}(3) protein models build their conditional paths from geodesics (Yim et al., arXiv 2023, Bose et al., ICLR 2024).

Any conditional path gives a valid marginal, but paths are not interchangeable in practice: diffusion and optimal-transport paths are both valid, and the second trains and samples faster (Lipman et al., ICLR 2023). It therefore matters whether the construction in use is the intended path, how far it departs when it is not (Theorem 1), and how much that can change training (ρ\rho, Part 3).

W2, continued. The policy experiment. §5.4 tests the criterion's prediction for a path it accepts: under the loss's own pullback metric the released path is already the geodesic, so changing it should bring no gain (lines 493 to 495). The comparator is the path a practitioner reading the metric as left-invariant would adopt, and it differs from the released path by 3.7 and 5.9 percent of the policy's translational and rotational error. Training on it changes the policy by less than 0.3 percent at a resolution of 0.10 to 0.26 percent (Table 3), and no layer moves further under the change of path than under a change of seed. We agree that Table 3 is an in-sample comparison, as Appendix D.9 states. On held-out episodes with three seeds the result is the same, and where the criterion rejects the path, training on a path closer to the geodesic lowers the held-out objective (Part 3).

W3 and Q2. The audit, separated by kind, and the unfrozen Euler-Lagrange residual. We agree that separating the three kinds of verdict is clearer, and Table 1 now names the kind in its metric column and caption. All four accepting verdicts hold without freezing, and the one violation is the same under the frozen and the unfrozen metric.

ConfigurationKind of verdictExact without freezing
EquiGraspFlowFixed metric, M=cIM = cIYes (§5.3, lines 422 to 424)
DiffDockFixed metric at each diffusion time, each block a multiple of the identity (Appendix D.1)Yes
FrameFlow, t<1/4t < 1/4Fixed-chart pullback, logarithm chart anchored at the data endpoint (Corollary 5, Part 1 (i))Yes, residual below
SmolVLAFixed-chart pullback (Lemma 16)Yes
FrameFlow, t>1/4t > 1/4Violation by the constant coupling block (Part 1 (iii))Same violation, 0.089 to 0.091 Å frozen and unfrozen
3D FlowMatchNo quadratic formNot applicable

The split itself, two acceptances from isotropy and two from a fixed chart, is a finding of the audit. None of these losses was designed for the criterion, and the paper explains why practical losses satisfy it: body-frame invariance forces isotropic blocks (Proposition 6), and a loss written in one fixed chart makes its straight path a geodesic (Lemma 16). The outcome is not evident without the criterion. Under the left-invariant reading SmolVLA's released path sits at least ten orders above the solver floor (§5.3, lines 427 to 429), and at fixed eigenvalues, turning the eigenbasis by 45 degrees moves a path from 1.8×10−131.8 \times 10^{-13} to 3.3×10−23.3 \times 10^{-2} Å (§5.2).

On Q2, the Euler-Lagrange residual of the varying metric below FrameFlow's gate vanishes to machine precision. We evaluated it for the metric of Part 1 (i), in exponential coordinates at the source, where neither the metric nor the path is trivial, with exact derivatives by automatic differentiation in double precision. Over 20 random endpoint pairs at each rotation amplitude from 29 to 172 degrees, which covers the medians of the five configurations with a Haar or anchored source, 117.1 to 132.4 degrees (§5.3), the largest residual relative to the momentum is 1.2×10−141.2 \times 10^{-14}, while perturbing the path by 1 percent of its amplitude raises it to 0.020 to 0.064. On 1 024 held-out states of real proteins, the released velocity target below the gate differs from the geodesic target of the unfrozen metric by 5.6×10−155.6 \times 10^{-15} of the network's regression residual (Part 3 (ii)).

AuthorsOpenReview ↗

Response to Reviewer QUTV (3/3): Q3, held-out episodes with three seeds, and the violating FrameFlow configuration trained

Q3. Held-out episodes, repeated seeds, and a configuration that violates the criterion. We ran both, and both outcomes are those the criterion predicts: no detectable effect where it accepts the released path, and an effect of the order of ρ\rho where it rejects it. We have added them to Table 3 and Appendices D.12 and D.13 of the revised PDF, marked in blue.

(i) Held-out episodes, three seeds. We retrained the two policies of Table 3, one on the released path and one on the exact geodesic of the frozen metric, using 1 525 training episodes and seeds 42, 43 and 44, and evaluated each on 1 000 chunks drawn uniformly from the frames of the 168 held-out episodes. The design of Table 3 is kept: sampling noise is shared, and intervals resample episodes. Entries are the change of the geodesic policy's error relative to the released policy, in percent, with paired 95 percent bootstrap intervals over episodes.

SeedPer-step trans.Per-step rot.Composed trans.Composed rot.MDE
420.001 [−0.066, 0.068]0.001 [−0.050, 0.054]−0.039 [−0.187, 0.116]−0.017 [−0.136, 0.113]0.075 to 0.219
430.045 [−0.022, 0.113]−0.003 [−0.056, 0.053]0.008 [−0.129, 0.150]−0.046 [−0.168, 0.077]0.078 to 0.200
440.015 [−0.066, 0.096]−0.002 [−0.059, 0.054]0.028 [−0.138, 0.193]−0.042 [−0.141, 0.058]0.081 to 0.239
Pooled0.021 [−0.029, 0.069]−0.001 [−0.041, 0.039]−0.001 [−0.116, 0.115]−0.035 [−0.122, 0.054]0.057 to 0.164

Every interval covers zero, and each pooled effect is smaller than the seed-to-seed standard deviation of the released policy on the same measure, 0.03 to 0.26 percent of its mean.

(ii) The violating configuration, measured on the released checkpoint. We built the above-gate metric without freezing, with the rotational pullback through Φ1\Phi_1, the translational term, and the auxiliary atom term at its released weight. On 1 000 pairs of §5.3 per seed, for seeds 42, 43 and 44, the released path departs from the geodesic by 0.089 to 0.091 Å (median per seed) both under the metric of §5.3, reproducing the 8.7 to 9.3×10−29.3 \times 10^{-2} Å reported there, and under the unfrozen metric. Removing the coupling block alone lowers this to 0.0025 Å, and setting the auxiliary weight to zero brings it to the solver floor. On the segments the training targets follow, from xtx_t to the data with t>1/4t > 1/4, the departure falls with the square of the remaining rotation, from 0.033 to 0.038 Å where 69 to 75 degrees remain to 0.0007 to 0.0018 Å where 10 to 16 degrees remain, as Theorem 1 predicts. Relative to the released network's regression residual on 1 024 held-out states, the departure of the velocity target is ρ=0.021\rho = 0.021 above the gate (median, interquartile range 0.008 to 0.051) and 5.6×10−155.6 \times 10^{-15} below it, with ρ\rho as defined in Table 5.

(iii) The violating configuration, trained. From the released checkpoint we fine-tuned FrameFlow on the released path and on a centred path, the factorized path of the template centred at its centroid (§5.1), which lies closer to the geodesic of the coupled metric on all 1 000 training segments (median 0.0065 against 0.0096 Å). Each run takes 2 300 optimizer steps on the released training set, under the released configuration with auxiliary weight one and the gate at t=1/4t = 1/4, for seeds 42, 43 and 44, and is evaluated on 64 SCOPe domains of 129 to 150 residues that are not in the training set. Entries are the change of each model's held-out training objective, centred minus released, in percent, with 95 percent bootstrap intervals over domains.

Seedt<1/4t < 1/4 (criterion accepts)t>1/4t > 1/4 (criterion rejects)
420.01 [−0.24, 0.27]−1.65 [−2.46, −0.92]
430.07 [−0.16, 0.30]−0.82 [−1.60, −0.15]
440.22 [−0.17, 0.60]−2.08 [−2.96, −1.23]
Pooled0.10 [−0.06, 0.27]−1.52 [−2.16, −0.95]

Each objective above is the model's own, so we also scored all six checkpoints on the same held-out states against two common targets: the released target, on which only the released-path model was trained, and the velocity of the solved geodesic of the unfrozen metric. Above the gate the centred-path model is closer to both, by 1.76 percent [−2.41,−1.19][-2.41, -1.19] in the training objective against the released target and by 0.51 percent [−0.77,−0.25][-0.77, -0.25] in distance to the geodesic velocity, pooled over the three seeds. Below the gate every difference covers zero.

Below the gate the two models do not differ detectably, and above it the path closer to the geodesic lowers the held-out objective on each of the three seeds, with intervals excluding zero, by an amount of the order of ρ\rho.

We hope that these results, together with the derivation and the audit in Parts 1 and 2, address all of the reviewer's questions, and we would be grateful if the reviewer would consider raising the rating in light of them.

Reviewer_fWH7OpenReview ↗
Rating
3
Confidence
1
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

I am unable to provide a reliable assessment because this paper is outside my area of expertise.

Critical strengths

I am unable to provide a reliable assessment because this paper is outside my area of expertise.

Critical weaknesses

I am unable to provide a reliable assessment because this paper is outside my area of expertise.

Questions for authors

NA.

1 comment

We thank the reviewer for the positive assessment, and for the opportunity to introduce the work briefly.

What we study. Flow-matching generative models now produce protein backbones, molecular poses and robot actions, all of which are rigid-body poses on SE(3)\mathrm{SE}(3). Such a model is trained along a conditional path that joins a source sample to a data sample, and the path its training loss calls for is the geodesic of the metric the loss induces.

The open problem. On SE(3)\mathrm{SE}(3) this geodesic generally has no closed form and requires a numerical boundary-value solve, so released models replace it with one of three closed-form constructions. How far these constructions lie from the geodesic had not been reported, and whether a given model needs a different path could so far be settled only by training two models and comparing them.

How we address it. (i) We derive the deviation in closed form: group interpolation departs from the geodesic at second order in the relative motion between the endpoints (Theorem 1). (ii) We prove a criterion, read off the training loss before any training, for when the deviation is exactly zero, and show that what decides it is the alignment of the metric with the rotation axis rather than its anisotropy (Propositions 4 and 6). (iii) Where the criterion fails, a closed-form correction removes one further order of the deviation per level, at a cost about 2 600 times below that of a numerical solve (Algorithm 1, Appendix D.6).

Impact. For existing models, the criterion gives a verdict from the loss alone: of six released configurations, four already train on the geodesic although none was designed for it, and the one violation is traced to an uncentred backbone template. For new models, it gives a design rule: a loss invariant under a change of body frame, with a block-diagonal Gram operator, makes the factorized path the geodesic for every endpoint pair (Proposition 6). Training confirms the criterion in both directions: where it accepts the released path, a different path brings no detectable change on held-out data, and where it rejects it, a path closer to the geodesic lowers the held-out objective (Table 3 and Appendices D.12 and D.13 of the revised PDF, with all changes marked in blue). Most of the theory also holds on other Lie groups (Appendix E).

We hope this overview is helpful, and we would be glad to answer any question.

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

The paper proposes a rigorous methodology for the evaluation of constructions used in place of the idealized conditional path. In particular, the authors introduce two `defect' quantities measuring distances of the constructions to the true solution, and between themselves, and use them to show that one can judge optimality rigorously. The paper then suggests an algorithm for correcting a non-optimal construction at two levels. Structurally, the paper begins with an introduction acting as a summary, and discusses preliminary mathematical concepts (Section 2) and its proposed methodological and theoretical contributions (Section 3). Then, related work (Section 4) is discussed as a bridge to computational results (Section 5) before concluding remarks (Section 6).

Critical strengths
  • This is an interesting and important problem for a subset of nonlinear flow matching; I am not in a position to check whether the non-reporting of the defect is accurate across a large body of literature. (But I suspect, from my own biased sample of relevant papers that I have read in the past, that it is a valid concern)
  • Section 2 is well-presented, and the introductory figures are very helpful towards supporting results.
  • The general methodology is rigorous and sound (although I have not checked the mathematics line-by-line), the mathematical framing is tight and supports the main results, and both the diagnostic and the correction it makes possible are real contributions for the group of interest (SE(3)).
  • To the extent of my understanding, the experiments support the adoption of the practical algorithmic tools both as diagnostic and computational, which has direct benefit to both pre-existing algorithms, and possibly ones to be further developed.
Critical weaknesses
  • My main concern is that the text is often hard to read, at times extremely hard to read, which I attribute to LLM writing. I am not faulting the practice of llm-drafting itself, just that I personally find this document hard to read, and unless it goes through a real re-draft, I do not see how it would be accessible to general readership (I had to use my knowledge of the mathematics to understand what the text was saying in section 3, it should be the other way around in general, especially in a short conference paper). I feel obliged to mention that I really enjoyed understanding the contributions once I was able to get past the text.
  • There are several place in the text where mathematics or 'technical-sounding' terms are introduced before they are defined. This makes the text hard to parse. For example, even in the abstract, I do know SE(3) because of my background, I do not know exactly what `endpoint amplitude' really means and it is a technical term introduced much later. The introduction is further riddled by such issues, for example τ\tau and uu are introduced without being defined.
  • The introduction is spent as a summary of the main text, which is not entirely helpful for an otherwise short paper. I find the general structure of the paper sound, Section 2 is the only one that reads cleanly (at least for me).
  • While the benefit of the proposed approach is presented numerically, the qualitative implication of that numerical benefit is not explained well.
Questions for authors
  • Some of the language surrounding the interpretation of the results is very strong, and it is unclear to me how much of the theory is true specifically for SE(3), and how much of it is generalizable to other groups with certain properties (and what those properties might be); The authors make the case, and I concur, that SE(3) is itself very important to be studied in its own right, but from a theoretical perspective, it is interesting to ask if some of the results of section 3 generalize?
  • What are the qualitative implications of applying the proposed methodology (especially the corrections) to the constructions of Section 5?
5 comments
AuthorsOpenReview ↗

Response to Reviewer NmqX (1/3): Overview, significance and novelty, and W1 to W3 on presentation

We thank the reviewer for a careful review, for finding the problem "interesting and important", the methodology "rigorous and sound", and the diagnostic and the correction "real contributions", and we are glad the reviewer "really enjoyed understanding the contributions". The weaknesses concern how the text reads and how the numbers are interpreted. None of them requires a change to a result, and we address each below.

In brief, the paper gives in closed form how far the conditional paths of SE(3)\mathrm{SE}(3) flow-matching models lie from the geodesic of the metric their loss induces, and from each other (Theorems 1 and 2, Proposition 3), decides from the loss before training when that distance is zero (Propositions 4 and 6, Corollary 5), and finds four of six released configurations already on it, with the one violation traced to an uncentred template and corrected in closed form (§5, Algorithm 1).

On whether the defect has been reported. The literature supports both the importance of the question and its novelty.

(i) The choice of path is known to matter. In Euclidean flow matching, optimal-transport paths train and sample faster than diffusion paths and generalise better (Lipman et al., ICLR 2023), couplings that simplify the flow make training more stable and inference faster (Tong et al., TMLR 2024), and the curvature of the generative trajectories is directly related to the truncation error of the solver (Lee et al., ICML 2023). On Lie groups and other manifolds the same method now generates protein backbones (Yim et al., arXiv 2023, Bose et al., ICLR 2024, Yue et al., ICML 2025), crystals (Miller et al., ICML 2024) and robot action chunks (Black et al., RSS 2025, Shukor et al., arXiv 2025), and generative models of residue frames already produce protein designs that have been characterised experimentally (Watson et al., Nature 2023). With the geodesic distance as premetric, Riemannian flow matching prescribes the geodesic as the conditional path (Chen and Lipman, ICLR 2024, Proposition 3.2), but for a general left-invariant metric this geodesic has no closed form and requires a boundary-value solve (Eq. 1), and no Riemannian metric on SE(3)\mathrm{SE}(3) has the screw motions of group interpolation as its geodesics (Žefran et al., IJRR 1999), so every implementation substitutes a closed-form construction.

(ii) The neighbouring literatures answer different questions (Appendix A). Work in kinematics establishes which metrics on SE(3)\mathrm{SE}(3) admit which optimal curves, but not how far a non-optimal curve lies from the optimal one (Žefran et al., IEEE TRA 1998, IJRR 1999, Appendix A.4). Rectification and optimal-transport couplings straighten the marginal field or change which endpoints are paired, and leave the curve joining a fixed pair unchanged (Liu et al., ICLR 2023, Pooladian et al., ICML 2023, Tong et al., TMLR 2024, Yue et al., ICML 2025, Appendix A.3). To our knowledge, no published comparison states how far the three constructions in use lie from each other or from the curve that minimises the conditional action (Appendix A.1). The audit covers every implementation we had identified by 22 September 2026 that trains a simulation-free generative model on SE(3)\mathrm{SE}(3) or one of its factors and releases both its weights and its training code, since the induced metric cannot be read from inference code alone (Appendix D.1, lines 2379 to 2383).

W1, W2 and W3. Presentation. We have revised the abstract, the introduction and the opening of Section 3, without changing any result, and the changes are marked in blue. The abstract and the introduction now define uu and τ\tau where they first appear, as Eq. 2, Table 5 and §3.1 do in the body, and the abstract states Theorem 1 in the quantity it is proved in, the relative amplitude ∥u∥\lVert u \rVert, the size of the rotation and translation that carry one endpoint to the other. The endpoint amplitude is ϵ=max⁡(∥ξ0∥,∥ξ1∥)\epsilon = \max(\lVert \xi_0 \rVert, \lVert \xi_1 \rVert), the larger of the norms of the two endpoints' logarithms (§3.1, lines 149 to 150). The introduction no longer restates Section 5: the two paragraphs that summarised its results are replaced by one that states what the results allow, namely that a released path is checked from its loss before any model is trained, and the results are stated once, in the list of contributions. The opening of Section 3 now ties each step of its argument to its subsection, the deviation (§3.1), the condition under which it vanishes (§3.2) and its correction (§3.3), and points to the notation collected in Table 5.

AuthorsOpenReview ↗

Response to Reviewer NmqX (2/3): W4 and Q2 on the qualitative implications of Section 5

W4 and Q2. The qualitative implications of the numbers, and of applying the corrections. For each released construction, the criterion states before training whether its conditional path should change and, where it should, what to change. Changing the path then affects a trained model detectably only where the criterion rejects it. These implications are stated across §5.3 and §5.4, and we collect them here.

(i) Implications for each released construction:

ConstructionVerdictQualitative implication
EquiGraspFlow, DiffDockIsotropic blocksThe released path already solves the geodesic equation for every pair, with the residual at the solver floor across sweeps up to 46 and 115 degrees (Table 2). The correction is identically zero, and the block weights change only the norm in which a deviation is reported (Appendix D.3).
FrameFlow, t<1/4t < 1/4Chart pullback (Corollary 5), aligned when frozenThe gap is 1.8×10−131.8 \times 10^{-13} Å, and rotating the same eigenvalues by 45 degrees raises it to 3.3×10−23.3 \times 10^{-2} Å (§5.2). The departure comes from misalignment, not from anisotropy.
SmolVLAFlat pullbackThe released path is exact for its own loss. A path that differs from it by 3.7 and 5.9 percent of the policy's translational and rotational error, composed over fifty actions, changes the trained policy by less than 0.3 percent (§5.4, Table 3), so the geometric size of a departure overstates its effect on the model by more than an order of magnitude.
FrameFlow, t>1/4t > 1/4ViolatedThe cause is a frame convention: the uncentred template couples rotation and translation, and centring it lowers the gap by a factor of 32 under the released metric (§5.1). If the loss is kept, the criterion selects the correction route: correcting in place lowers the gap by a factor of 2.8, while switching to group interpolation and correcting raises it by a factor of 14.6 (§5.3).
3D FlowMatchNo quadratic formThe criterion does not apply (§5.1).

(ii) In task units, the single violation is measurable but well within tolerance, and the source distribution fixes in advance which tool applies. FrameFlow's departure above its gate, 8.7 to 9.3×10−29.3 \times 10^{-2} Å on 3 000 pairs (§5.3), is 4.4 to 4.7 percent of the 2.0 Å self-consistency RMSD below which FrameFlow calls a backbone designable (Yim et al., arXiv 2023), and the action head's composed deviation, 4.5 mm over fifty actions, is 0.149 of the 30 mm LIBERO placement tolerance (Appendix D.7). The operating amplitude is set by the source distribution (Proposition 20): the five configurations with a Haar or anchored source sit at median relative rotations of 117.1 to 132.4 degrees, where the exact criterion decides, and the action head at 3.46 degrees, where under the frozen metric one level of correction lowers the gap by a factor of 125 (§5.3).

(iii) In training, changing the path has no detectable effect where the criterion accepts it, and improves the model where the criterion rejects it. On the action head we trained four policies that differ only in the conditional path, the released construction, one and two levels of correction, and the exact geodesic of the frozen metric. They differ by less than 0.3 percent (Table 3), and the parameters of the once-corrected policy move by 0.46 of the change a different seed causes, at the 90th percentile over all tensors (Table 15). Repeating the comparison of the released path with the frozen-metric geodesic on held-out episodes with three seeds, every interval again covers zero at a resolution of 0.057 to 0.164 percent. On FrameFlow above its gate, where the criterion rejects the path, we fine-tuned the released checkpoint with its training configuration unchanged, auxiliary weight one and the gate at t=1/4t = 1/4, for 2 300 steps with seeds 42, 43 and 44 on the released path and on a centred path that lies closer to the geodesic on all of 1 000 training segments. Above the gate the centred path lowers the held-out objective by 0.82 to 2.08 percent on each seed, with 95 percent intervals excluding zero, and below the gate it leaves the objective unchanged within error. Above the gate and on identical held-out states, it is also closer to the released target the other model was trained on, by 1.76 percent, and to the velocity of the solved geodesic, by 0.51 percent.

Changing the path therefore makes a detectable difference to a trained model only where the criterion rejects it, and then by an amount of the order of ρ\rho, the ratio of the per-increment departure to the network's regression residual (Table 5), which is 0.021 for FrameFlow above its gate and is computed from the released checkpoint, before any alternative path is trained. We have added this conclusion at the end of Section 5 (lines 496 to 505), and Table 3 now reports both held-out comparisons.

AuthorsOpenReview ↗

Response to Reviewer NmqX (3/3): Q1 on the strength of the language, and on which results extend beyond SE(3)

Q1, first part. The strength of the language. We have qualified each strong claim with the condition under which it holds. In the contributions and the conclusion, "already optimal" is replaced by the statement that these configurations already train on the geodesic of the metric their loss induces, which is what Table 2 verifies. What is decided before training is whether the path departs from that geodesic, as the abstract states, and §5.4 now gives the effect of a departure on a trained model through ρ\rho (Table 5).

Q1, second part. Which results are specific to SE(3)\mathrm{SE}(3), and what properties do the others need? Most of the theory already holds on any finite-dimensional Lie group with a left-invariant metric, and the two results specific to SE(3)\mathrm{SE}(3), the alignment criterion and the body-frame result, extend under the conditions below, which we derived for this response and have added as Appendix E.

(i) Theorems 1 and 2, Propositions 7 and 15, Lemma 16 and the composition formula of Appendix C.14 hold on any finite-dimensional Lie group (Table 7, and the paragraph "Generality" of Appendix C.12), and Theorem 2, proved in the free Lie algebra and checked in exact arithmetic (Appendix C.3), holds on every Lie algebra within the convergence domain of the Baker-Campbell-Hausdorff series. Propositions 3, 4 and 6 and Corollary 5 are stated for SE(3)\mathrm{SE}(3). We checked the general statements numerically on SO(4)\mathrm{SO}(4) with an anisotropic metric and on SL(2,R)\mathrm{SL}(2,\mathbb{R}), which admits no bi-invariant Riemannian metric (Milnor, Adv. Math. 1976, Lemmas 7.4 and 7.5): the profile of Theorem 1 is recovered against the solved geodesic, and each correction level of Proposition 7 removes one order, with final local slopes of 2.00, 3.01 and 4.00 on SO(4)\mathrm{SO}(4) and 2.00, 2.95 and 4.00 on SL(2,R)\mathrm{SL}(2,\mathbb{R}).

(ii) The alignment criterion extends to every compact connected group, where alignment means commutation. A one-parameter subgroup has constant body velocity, so by Eq. 1 it is a geodesic exactly when Λ(u)=0\Lambda(u) = 0. The Lie algebra of a compact connected group carries an Ad-invariant inner product ⟨⋅,⋅⟩0\langle \cdot, \cdot \rangle_0 (Milnor, Adv. Math. 1976, Lemmas 7.4 and 7.5). Writing M=⟨S⋅,⋅⟩0M = \langle S \cdot, \cdot \rangle_0 with SS symmetric and positive, invariance gives ⟨Su,[u,y]⟩0=⟨[Su,u],y⟩0\langle S u, [u, y] \rangle_0 = \langle [S u, u], y \rangle_0 for every yy, so Λ(u)=0\Lambda(u) = 0 exactly when [u,Su]=0[u, S u] = 0. On so(3)\mathfrak{so}(3) this says that uu is an eigenvector of SS, which is Proposition 4 and the classical condition for steady rotation of a free rigid body about a principal axis (Arnold, Mathematical Methods of Classical Mechanics, 1989, Chapter 6), and on SE(n)\mathrm{SE}(n) with M=diag(J,βI)M = \mathrm{diag}(J, \beta I) the factorized path is a geodesic exactly when [ω,Sω]=0[\omega, S\omega] = 0, since the translational part satisfies the geodesic equation identically, as in the proof of Proposition 4. In higher rank, commutation is weaker than the eigenvector condition. On SO(4)\mathrm{SO}(4), u=E12+E34u = E_{12} + E_{34} is not an eigenvector of SS but commutes with SuS u, and departs from the solved geodesic by less than 3.5×10−153.5 \times 10^{-15}, while a generic direction departs at the second order of Theorem 1. This is consistent with a known result for the multidimensional free rigid body, whose steady rotations need not be about principal axes when eigenvalues of the angular velocity coincide (Izosimov, J. Phys. A 2012).

(iii) The body-frame result extends to SE(n)\mathrm{SE}(n), and the property it relies on is irreducibility. Invariance of the loss under every rotation of the body frame makes the rotational block Ad-invariant, hence bi-invariant, so every rotational one-parameter subgroup is a geodesic (Appendix C.4). What remains is that the coupling block vanish. It is an SO(n)\mathrm{SO}(n)-equivariant map between the representations on so(n)\mathfrak{so}(n) and on Rn\mathbb{R}^n, so by Schur's lemma (Fulton and Harris, Representation Theory, 1991, Lemma 1.7) it vanishes whenever the two share no irreducible component, which fails only for n=3n = 3, where so(3)≅R3\mathfrak{so}(3) \cong \mathbb{R}^3. Solving for all invariant metrics for nn from 2 to 6 confirms this. For every nn other than 3, body-frame invariance therefore suffices on its own, and SE(3)\mathrm{SE}(3) is the one case that needs the block-diagonal condition, which is why Proposition 6 states it.

We hope that the revision and these answers address all of the reviewer's concerns, and we would be grateful if the reviewer would consider raising the rating in light of them.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of the rebuttal and the revision for the Program Chairs, Senior Area Chairs, Area Chairs and Reviewers

We thank the Program Chairs, the Senior Area Chairs, the Area Chairs and the four reviewers for their time and careful reading. As the discussion period draws to a close, we summarise our responses in one place.

We have answered every weakness and every question of each reviewer, point by point, in dedicated replies: W1 to W4, Q1 and Q2 of NmqX, W1 to W3 and Q1 to Q3 of QUTV, and the weakness of gh1b with each of its questions (better trained, more stable, faster to converge, and a new interpolation scheme). For these answers we derived the metric of every audited configuration, ran new experiments with three seeds, extended Appendix C.8 and added Appendices D.12, D.13 and E. The revised PDF marks every change in blue, and no result of the submission changes. The main additions are as follows.

(i) The metric derived from the training objective, and the audit by kind (QUTV W1, W3, Q1, Q2). For each audited configuration we derived the metric that the training objective induces on the evolving state. Every verdict of Table 1 holds without freezing, Table 1 now states whether each rests on a fixed metric or on a fixed-chart pullback, and below FrameFlow's gate the Euler-Lagrange residual of the unfrozen metric is at most 1.2×10−141.2 \times 10^{-14} from 29 to 172 degrees (Appendix C.8).

(ii) Held-out data and repeated seeds (QUTV W2, Q3, gh1b W1). On held-out episodes with seeds 42, 43 and 44, every interval of the action-head comparison covers zero at a resolution of 0.057 to 0.164 percent, and the two policies do not differ in seed spread, convergence or closed-loop success on LIBERO-Spatial (Table 3a, Appendix D.12).

(iii) A configuration the criterion rejects, trained (QUTV Q3, gh1b W1, NmqX Q2). Fine-tuning FrameFlow above its gate on a path closer to the geodesic lowers the held-out objective by 1.52 percent [−2.16,−0.95][-2.16, -0.95] pooled over three seeds, with every seed's interval excluding zero, and makes no detectable difference below the gate. The ratio ρ=0.021\rho = 0.021, read from the released checkpoint before any alternative path is trained, gives the scale of the effect (Table 3b, Appendix D.13).

(iv) Actionable consequences and the interpolation scheme (gh1b W1, NmqX W4, Q2). For each released construction, the criterion states before training whether its path should change and, where it should, what to change. Algorithm 1 removes one order of the gap per level at a cost 2 601 times below a boundary-value solve at the median (Appendix D.6), and Proposition 6 gives a design rule for new losses, in which alignment rather than anisotropy decides optimality (§5.2).

(v) The premise (QUTV W2). The abstract and the introduction now state why the conditional path of least kinetic energy is the reference: among Gaussian conditional paths the conditional optimal-transport path minimises the conditional kinetic energy, an upper bound on that of the marginal path (Shaul et al., ICML 2023), and Appendix C.1 shows that the bound holds on any Riemannian manifold.

(vi) Generality beyond SE(3) (NmqX Q1). Most of the theory holds on any finite-dimensional Lie group. The alignment criterion extends to every compact connected group as the commutation [u,Su]=0[u, Su] = 0, and the body-frame result extends to SE(n)\mathrm{SE}(n), where the block-diagonal condition is needed only for n=3n = 3. Both are checked numerically on SO(4)\mathrm{SO}(4) and SL(2,R)\mathrm{SL}(2,\mathbb{R}) (new Appendix E).

(vii) Clarity and the strength of the claims (NmqX W1 to W3, Q1). The abstract and the introduction define uu and τ\tau where they first appear, the introduction no longer restates Section 5, and the opening of Section 3 ties each step to its subsection. "Already optimal" is replaced by the statement that these configurations already train on the geodesic of the metric their loss induces.

(viii) Novelty relative to prior work (NmqX). Prior work shows that the choice of conditional path matters in flow matching, while work in kinematics and on rectification addresses different questions, so to our knowledge no published comparison states how far the three constructions in use lie from the geodesic (Appendix A).

We would be very grateful if the reviewers could let us know whether our replies address their concerns before the discussion period ends, and we remain glad to answer any further question.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.