acceptodds
Under review as a conference paper at ICLR 2027

Endpoint Singularity and Temporal Heterogeneity in Flow-Matching Transport

Abstract

Flow matching learns a time-dependent velocity field transporting a source distribution toward the data distribution. We study the endpoint geometry of this transport. A lower-dimensional target forces the endpoint flow map to lose rank. In a quadratic Hamilton–Jacobi formulation, this appears through a Riccati equation for the velocity Jacobian whose critical eigenvalue corresponds to endpoint rank loss. We further confirm this criticality empirically: the divergence of a single trained model grows sharply as . Fine-tuning modifies the learned velocity field and can be viewed as a perturbation of it. We measure the temporal profile of this perturbation and find that it is non-uniform in flow time. We use this structure to define a peak-based query distribution for dynamics-level watermark detection, in which the query budget is concentrated on flow times where the watermark signal has larger energy. On watermarked flow-matching models, the peak-based distribution is more robust to fine-tuning than uniform and fixed-mixture sampling, retaining higher detection accuracy at intermediate fine-tuning steps.

Primary area:
generative models
Keywords:
flow matching

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_wt9sOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies endpoint geometry in flow matching. Within a Hamilton–Jacobi framework, it uses Riccati dynamics of the velocity Jacobian to analyze rank loss of the flow map and divergence singularities, with an OT-paired toy example checking the theory. The authors also measure temporal variation in fine-tuning perturbations on MNIST and CIFAR-10 and propose a watermark query strategy that measures readout power and concentrates queries around two peaks. At some fine-tuning stages, this strategy recovers watermark messages more accurately than uniform sampling and a fixed two-peak mixture.

Critical strengths
  1. The endpoint analysis has value, and the toy experiment fits its assumptions. The paper distinguishes rank loss of the flow map from changes in divergence and gives an explicit critical condition under the Hamilton–Jacobi assumptions. The toy experiment uses exact OT pairing and checks both singular values and divergence, providing a direct comparison with the idealized analysis.

  2. The query strategy does show gains over the baselines. The authors compare against both uniform sampling and a fixed two-peak Gaussian mixture. At the same online query budget, accuracy on CIFAR-10 after 300 fine-tuning steps is 0.90 for peak sampling, 0.55 for the fixed mixture, and 0.35 for uniform sampling. This difference is worth investigating and suggests that choosing query times can help.

Critical weaknesses

1. I do not see how the theory supports the proposed method, and the method itself adds relatively little.

The abstract connects endpoint criticality, temporal variation in fine-tuning perturbations, and watermark query allocation, but the paper does not establish the key links. A sharp change in divergence near the endpoint alone does not confirm the specific critical-eigenvalue condition. The Riccati analysis assumes Hamilton–Jacobi characteristics, but the MNIST and CIFAR-10 models are trained using independent coupling. The paper does not establish that the same dynamics apply to this training setup.

Nonuniform fine-tuning perturbations also do not by themselves explain the watermark readout power or where to query. Section 5 uses a measured power profile, not the geometric quantities analyzed earlier; the power-based strategy would still make sense without the endpoint analysis. With the embedding and decoder fixed, the main addition is adjusting the locations and weights of two query peaks. The peaks in Figure 6 largely follow the imposed sinusoidal envelope. The paper does not isolate the contributions of locations and weights, report the extra calibration cost, or test a different envelope. The evidence therefore supports improved querying for this watermark, without showing what additional value the endpoint theory brings to it.

2. The detection protocol is unclear, and the experiments mainly support message recovery on watermarked models.

Algorithm 1 lists the codeword cmc_m as an input, while Eq. (34) defines the decoded message m^\hat m through an argmax over candidate codewords. It is unclear whether the task is known-message verification or unknown-message recovery. The scalar readouts in Eqs. (30) and (33) versus the vector correlation in Eq. (34) also need explanation. These issues directly affect how the actual detection procedure should be understood.

Figure 7 reports message-recovery accuracy on watermarked models, without unwatermarked-model or wrong-key verification controls and false-positive measurements. Knowing one's own key and message is a legitimate setting. However, if ownership verification is an intended application, further evidence is needed to show that models without the target watermark are reliably rejected. The current results support improved message recovery in the tested setting, but do not by themselves establish that verification capability.

Questions for authors
  1. Why does the Hamilton–Jacobi/Riccati analysis apply to the independent-coupling experiments? Beyond qualitative temporal variation, what testable prediction does the endpoint analysis make about query locations or detection gains?

  2. At the same online budget, can you separately compare adaptive locations with equal weights and fixed locations with adaptive weights, and report the extra calibration cost and reuse conditions? Do gains persist with a different envelope, or should the conclusion be restricted to the current watermark setup?

  3. Does the detector know the target message in advance, and what readout and decoding equations are actually used? Please specify the task evaluated in Figure 7. If ownership verification is intended, please provide unwatermarked-model and wrong-key controls and detection rates at fixed false-positive rates after independent threshold calibration. Otherwise, please state explicitly that the results only cover message recovery.

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

This paper studies endpoint singularity and temporal heterogeneity in flow matching. The authors first analyze the dimensional collapse of flow maps when transporting a full-dimensional source distribution to a lower-dimensional target, using a Hamilton–Jacobi formulation and Riccati dynamics of the velocity Jacobian. They then empirically examine the temporal profiles of velocity divergence and fine-tuning-induced perturbations on a toy problem, MNIST, and CIFAR-10. Finally, the authors propose a peak-based query sampling strategy for dynamics-level watermark detection, reporting improved robustness to fine-tuning compared with uniform and fixed-mixture sampling.

Critical strengths

The analysis of endpoint dimensional collapse provides a potentially interesting geometric perspective on flow matching.

Critical weaknesses
  1. Unclear problem formulation and presentation. I found the paper difficult to follow despite careful reading. In particular, the paper does not clearly establish a central research question or a coherent narrative connecting its different components. It moves from a theoretical analysis of endpoint singularity to empirical observations of fine-tuning perturbations and then to a watermark detection algorithm. However, the connections between these parts are insufficiently explained. Consequently, it is difficult to determine whether the main contribution is intended to be a theoretical understanding of flow matching dynamics or a new watermark detection method.

  2. Insufficient connection between the theoretical analysis and the proposed method. The theoretical analysis is developed under a quadratic optimal-transport formulation, whereas the MNIST and CIFAR-10 experiments use flow matching models trained with independent coupling. The paper does not establish how the theoretical results directly apply to these experimental settings. More importantly, the connection between endpoint singularity and the proposed watermark detection method is unclear. The watermark uses a manually specified sinusoidal temporal envelope, which already induces strong temporal variation in the watermark signal. Therefore, the improvements from peak-based querying do not necessarily follow from the endpoint singularity analysis. Overall, I do not find convincing evidence that the theoretical findings are essential to or directly justify the proposed method.

Questions for authors

See weakness

Reviewer_vvwQOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies endpoint singularities and temporal heterogeneity in flow-matching transport. It relates dimensional collapse to the behavior of the velocity Jacobian under a Hamilton–Jacobi formulation, and examines time-dependent velocity and divergence profiles in trained models. Motivated by these observations, it proposes a peak-based query distribution for watermark detection. Experiments on synthetic data, MNIST, and CIFAR-10 illustrate the analysis and evaluate detection robustness under fine-tuning.

Critical strengths
  • An interesting geometric perspective. The paper connects transport toward a lower-dimensional target with rank loss in the endpoint flow map. Examining this connection through the velocity Jacobian provides a concrete perspective on transport behavior near the endpoint.

  • A useful analytical example. The toy problem has an explicit transport solution, allowing the flow-map deformation and divergence to be examined directly. It provides a concrete illustration of the endpoint-collapse mechanism under the theoretical setting considered.

Critical weaknesses
  • Insufficient support for the central claims. The paper does not establish that its theoretical analysis applies to the flow-matching models used in the experiments. Inconsistencies in the SNR and watermark decoding definitions further make the reported results difficult to verify.

  • Inadequate presentation of the empirical evidence. The main text makes claims about endpoint behavior and watermark robustness but contains no result figures or tables. It provides only brief descriptions and selected numerical results, with all empirical figures and tables deferred to the appendix. This makes the empirical argument difficult to assess from the main paper.

  • Missing required AI use statement. The submitted manuscript does not include the AI use statement required by the ICLR 2027.

Questions for authors

See weaknesses.

1 comment

The rebuttal does not change my view; I firmly stand by my score of 1.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.