- 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
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, , and 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:
| Seed | Released path | Geodesic of the frozen metric | McNemar |
|---|---|---|---|
| 42 | 30.6 [26.7, 34.8] | 29.8 [26.0, 34.0] | 0.74 |
| 43 | 33.4 [29.4, 37.6] | 32.0 [28.1, 36.2] | 0.49 |
| 44 | 32.4 [28.4, 36.6] | 30.4 [26.5, 34.6] | 0.34 |
| Pooled | 32.1 [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 of the network's regression residual, against 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 , 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.
| Seed | (criterion accepts) | (criterion rejects) |
|---|---|---|
| 42 | 0.01 [−0.24, 0.27] | −1.65 [−2.46, −0.92] |
| 43 | 0.07 [−0.16, 0.30] | −0.82 [−1.60, −0.15] |
| 44 | 0.22 [−0.17, 0.60] | −2.08 [−2.96, −1.23] |
| Pooled | 0.10 [−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 in the training objective, and to the velocity of the solved geodesic, by 0.51 percent . 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 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.
- 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.
- 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.
- 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.
- 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.
- 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.
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) 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) 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 1.8×10−13 to 3.3×10−2 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).