- 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
-
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.
-
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 as an input, while Eq. (34) defines the decoded message 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
-
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?
-
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?
-
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.
The rebuttal does not change my view; I firmly stand by my score of 1.