acceptodds
Withdrawn from ICLR 2027

Straightness and Intrinsic Geometry in OT Flow Matching

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Optimal transport (OT) flow matching promotes straight ambient trajectories, but exact ambient fitting need not recover intrinsic transport velocities on curved supports. We quantify this mismatch and examine whether tangent-projected straightness can detect it. For short circle arcs and full-dimensional product spherical bands, we prove shared ambient and intrinsic optimality and derive exact midpoint velocity-gap laws. The gap determines the leading excess kinetic action of the projected path, despite matching endpoint laws. A structured product-torus extension accommodates correlated source densities. Exact calibrations show that the diagnostic can vanish despite a positive intrinsic gap, while nonlinear representations can produce a positive score on a flat path. Separate CIFAR-10 experiments evaluate training behavior and sample selection without intrinsic ground truth. Across two scoring representations and three quality metrics, lower-score retention performs worse on average than matched random retention; reversing DINOv2 rankings yields metric-dependent gains. Projector ablations preserve this selection pattern. An exact finite-pool MMD identity explains why favorable bucket association need not improve the retained distribution.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
flow matching, optimal transport, manifold geometry, trajectory straightness, geometric diagnostics

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_GenXOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies a mismatch between ambient OT flow matching and intrinsic transport on curved embedded manifolds. For short circle arcs, product spherical bands, and a separable product-torus construction, it proves that ambient and intrinsic squared-cost OT can share the same endpoint coupling while their midpoint velocity fields differ. It also relates this field gap to the excess intrinsic kinetic action of the nearest-point projection of the ambient chord. The paper then asks whether a tangent-projected feature endpoint-chord residual can diagnose this mismatch. Exact calibrations show two complementary failures: the residual can vanish in the presence of a positive intrinsic field gap, and nonlinear features can make it positive on an ambiently straight path. Finally, CIFAR-10 experiments examine the score during training and as a sample-selection rule; the predeclared lower-score rule performs worse than matched random retention, while reversed DINOv2 rankings yield metric-dependent gains. An exact finite-pool MMD identity explains why favorable separation between score buckets need not imply improvement over random retention.

Critical strengths
  1. Careful isolation of a genuine geometric mismatch. The paper does more than compare an ambient chord with a geodesic pairwise: for its structured examples it proves common ambient/intrinsic optimality, midpoint-law agreement, and pair identification, thereby preventing conditional averaging or a coupling change from explaining away the midpoint field gap. The projected-path action calculation is also a clean consequence and clarifies that correct endpoint laws do not imply minimal intrinsic action.

  2. Unusually transparent negative calibration and evaluation. The paper openly reports that the proposed residual cannot certify intrinsic error and that the predeclared lower-score selection direction fails. It separates bucket association from improvement over matched random retention and derives an exact MMD accounting identity. This is more rigorous and informative than reporting a favorable post-hoc ranking alone.

Critical weaknesses
  1. The strongest field-level theory is too restricted to support the broad significance suggested by the paper. The exact conditional-field and action results require short common arcs, product spherical bands with a shared independent directional law, or coordinatewise monotone maps with separable costs. The more general curvature expansion is explicitly only a same-pair statement and does not establish a common OT coupling or a field law. Moreover, the reported excess action is only 0.074% at the learned controls' shift of 0.6 and 1.29% at shift 1.2. Thus the paper establishes a precise phenomenon in constructed settings, but not that it is consequential for realistic coupled high-dimensional transports. A substantially stronger paper would either extend the result to a nonseparable coupled setting or demonstrate a practically meaningful consequence against an intrinsic/geometric flow-matching baseline.

  2. The feature-probe contribution is conceptually weak once its calibration failures are established. For affine features, the zero-score result follows because the feature derivative along an exact linear path equals its endpoint chord before projection; zero-preserving post-processing cannot recover information that is absent from this residual. This is a valid negative result, but it excludes only a narrow comparator family and does not yield a replacement diagnostic. The CIFAR-10 section then studies a heuristic already shown not to measure the geometric target, without intrinsic ground truth. Consequently, the paper's theory and image experiments do not form a convincing validation chain: the experiments measure representation-specific behavior and selection utility, not the intrinsic mismatch developed in Sections 3-4.

  3. The image evidence is insufficient for general conclusions about sample selection. The primary selection study uses only three fixed CIFAR-10 models; the confirmation pool changes candidate noise but reuses the same models and real references. The reversed-DINO effect changes sign across feature metrics, and there is no transfer study across architectures or datasets. The reported FID values are also uniformly high (approximately 37-49), including a random-half baseline near 39. This does not by itself indicate an implementation error: each FID estimate uses only 1,024 generated images against the 10,000-image test set, rather than the conventional 50,000-sample CIFAR-10 protocol, and the paper correctly notes that finite-sample FID is biased. It does, however, make the small reversed-DINO improvements difficult to interpret. On the discovery pool, the apparent gain over random is only about 1.4-1.6 FID points; on the confirmation pool the mean gain is 1.27 and occurs in only two of three models. These values should not be compared with published FID-50k results or treated as robust evidence of improved image quality. The paper also calculates that 50% forecast retention requires 76 imagewise velocity evaluations per retained output versus 50 for direct generation, before feature/PCA costs. Overall, the experiments support a narrow negative statement about the tested score and checkpoints, but do not establish a generally useful or generally harmful selector. An image-like benchmark with known latent manifold geometry, plus independent-model and cross-dataset replication, is needed to bridge the geometric and empirical parts.

Questions for authors
  1. Can the common-coupling field result be extended to any genuinely nonseparable transport, or can the authors provide evidence that the measured action gap materially affects learning or sampling relative to Riemannian/metric flow matching?
  2. What conclusion about the feature probe is intended to generalize beyond the exact comparator definition and the three fixed CIFAR-10 checkpoints? Which result would falsify that broader conclusion?
  3. Can the authors add a benchmark with image-like observations but known latent manifold geometry, so that the proposed score can be compared directly with the true intrinsic velocity gap and projected-path excess action?
Reviewer_o1twOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies whether ambient straight-line OT flow matching recovers intrinsic transport velocities when data lie on a curved manifold. For short circle arcs, product spherical bands, and coordinatewise product-torus transports, it proves that the ambient and intrinsic OT couplings coincide and derives exact midpoint velocity-gap laws of order d³/(24τ²). It also relates the gap to the excess intrinsic kinetic action of the projected path. The paper then proposes a feature-space, tangent-projected straightness probe, and shows by construction that the probe can vanish despite a positive intrinsic gap and can be positive on a flat path. Finally, it tests the probe for selecting generated CIFAR-10 samples. Lower-score retention is worse than matched random retention, and reversed DINO rankings give metric-dependent gains. A finite-pool MMD identity explains why a favorable bucket gap need not improve the retained distribution.

Critical strengths
  • The exact results within the stated classes (Theorems 1–2, Corollary 3) are careful and numerically verified.
  • Reporting is transparent: directions are predeclared, confirmatory and post-hoc analyses are separated, and per-seed values are released.
  • The finite-pool MMD identity (Eq. 10) is a simple, reusable criterion for evaluating subset selection.
Critical weaknesses
  • The paper has no clear positive contribution. Its own constructions show that the proposed probe fails as a measure of intrinsic error, and lower-score retention loses to random retention in every seed (Table 1).
  • The theory covers only narrow structured classes, and the effect is tiny (excess action of 0.074% at δ=0.6). Its relevance to realistic generative modeling is not argued.
  • The presentation is very hard to follow, with heavy qualifications, private terminology, and confirmatory, post-hoc, and audit analyses mixed without a clear central claim.
Questions for authors
  1. What is the single main takeaway the authors intend for flow-matching practitioners?
  2. Is there any realistic setting in which the intrinsic–ambient gap measurably affects sample quality or a downstream metric?
Reviewer_saFXOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper shows OT flow matching only minimises the transport in ambient space and not in intrinsic manifolds. For cases such as circular arcs, spherical bands, or torus, it showed the ambient and intrinsic OT share the sample coupling and midpoint law, but the tangent-projected ambient field differs from the intrinsic field, linking it to excessive intrinsic kinetic action. Also proposing a tangent projected endpoint-chord diagnostic, which can be zero despite a positive intrinsic gap.

Critical strengths

The paper correctly discussed how fitting on ambient space does not lead to intrinsic optimality. Also, tying the same to excess kinetic action.

Critical weaknesses
  1. I believe one of the biggest weaknesses of this work is it's writing, which is very convoluted and hard to understand, without several passes. I urge the authors to rewrite the entire paper to that effect.
  2. If I understand correctly, the paper assumes a coupling on Manifold M x M. However, the most common formulations of flow matching use a Gaussian source.
  3. The evidence is only shown on one real dataset, i.e. CIFAR-10.
  4. The theorems apply to cases (circles, spheres, toruses), where geometrical flow matching already exists and handles the intrinsic targets well. The paper does not discusses any setting, where the gap changes anything for practical decision making.
  5. The gap is tiny (0.074% excess action at delta = 0.6), and no evidence of whether it changes the training, or sample quality,
Questions for authors

See weaknesses above

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

Related papers

Discussion (0)

Sign in to comment.