- Rating
- 3
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 3
Read review
Paper summary
This paper introduces Characteristic-Steered Flow Matching for constraints accessible through a binary membership oracle. It pairs invalid samples with certified feasible anchors, then transports the endpoint corrections backward through a reference flow to construct velocity-training targets. An optional runtime repair handles remaining violations. Experiments on synthetic sets, constrained control, and trajectory planning examine whether this improves validity while preserving coverage and useful generative structure.
Critical strengths
The paper aims to solve the discrete membership problem into a constraint supervision along generation path which solves a timely question. The stengths are:
- Pairing rejected samples with feasible anchors gives a concrete correction, and the reverse-flow construction provides a clear first-order justification for distributing it along the trajectory. This is a neat way to use an oracle without requiring constraint gradients.
- The paper shows good comparison over existing method: the dimension-scaling results show cases where CSFM maintains coverage while FM-RE collapses. The single-round planning results are also encouraging: CSFM achieves high safety and success at substantially lower latency than HardFlow.
Critical weaknesses
I think a further ablation will help: Table 15 ablates the steering loss, correspondence choice, and correction strength, but does not compare the reversed flow transported r against simply using r with the same anchors and training setup. Continued flow-matching training already gives similar validity in several ablation cases. It therefore remains unclear how much of the improvement specifically requires the paper’s central transport mechanism, rather than a gentler anchor-based correction.
The planning results leave me somewhat unsure when the extra adaptation is worth it compared with inference-time search and repair. In Table 1, Base + CEM already matches CSFM + CEM on avoidance safety and success and achieves higher mean safety and success on the maze task at similar latency. CSFM does show a clear benefit in the single-round setting, so I am not suggesting adaptation is unnecessary. However, a search-budget sweep would help show where this benefit holds and how much inference effort the extra training actually saves.
Questions for authors
- Could you compare CSFM against an otherwise identical correction baseline i.e. using A_r be identity matrix, keeping the anchors, correction strength, and training budget fixed? This would directly test whether characteristic transport is needed.
- Could you provide a candidate/search-budget sweep beyond the existing single-round and CEM settings, comparing CSFM and the base model under matched inference costs?
- Relatedly, it would be useful to discuss Physics-Constrained Flow Matching (PCFM), which applies constraint corrections during inference without retraining the generator. This raises a broader question about learning corrections versus applying them during sampling.
4 comments
Thank you again for your review and your suggestions. We would like to know whether the identity control and budget sweep address your questions, and we are happy to answer any remaining ones.
I thank the authors for extra experiments and results. The identity ablation is what I was looking for and the budget sweep also answers my question about CEM. The explanation of PCFM makes sense. Given the different constraint information it requires, I’m fine with discussing it as related work without a direct comparison. However, though the method shows advantage with limited inference time, overall evidence still seems mixed to me to rate as a strong accept and hence I will maintain my weak accept evaluation.
We agree that the evidence is mixed: CSFM helps most where the base fails or inference time is limited, and the final version can state this scope more clearly. Thank you for your support.
We thank the reviewer for the positive assessment and for precise, actionable remarks: the identity control and the search-budget sweep are now in the paper. To improve our presentation, the revised PDF defines both Suite C deployments (Sec. 4.3) and provides their per-replan oracle budgets (App. C.2, Table R7), expands metric abbreviations where they are introduced, and states which correspondence rule each suite uses (Sec. 3.4).
W1 / Q1. CSFM vs. an otherwise identical correction with A_t = I.
All three reviewers asked for ablations that isolate the transport. Yours, the same correction with A_t = I and identical anchors, alpha, and budget, is now in the paper. Validity, seeds matched across arms (all cells: Table R2):
Transport adds validity on 10 of 13 Suite A cells, most where the base fails most. On three Solid Torus cells and on both Suite B control tasks, the reviewer's "gentler anchor-based correction" indeed suffices: the two tie, and identity is often more uniform. On continued FM: where the base fails, as on Box d=32, CSFM gains clearly, and it is never below continued FM in validity or coverage on any analytic cell (General Response, point 2). The ablations requested by the other reviewers are in point 1 of the General Response.
W2 / Q2. When is adaptation worth it vs. inference-time search, at matched inference cost?
The sweep below makes explicit a benefit we had under-reported: inference time. With a good base, CEM search is very strong, but slow. To reach CSFM's single-round outcome, the base needs CEM, at about 3 times the time per replan in avoidance and twice in maze. We fully concede the maze row of Table 1: with CEM, the base is higher on both safety and success at similar latency. Adaptation therefore pays when search is limited or latency-bound; this is what we expect: removing edge cases from the flow will reduce expressivity, and our goal is to do that minimally while reducing errors. Oracle calls and time per replan for every point (Table R7, Fig. R-B):
Avoidance
Maze
(same evaluation seeds for all methods, 50 episodes per seed.)
Q3. PCFM: learning corrections vs. applying them during sampling.
Thank you for the pointer; we have added PCFM (Utkarsh et al., NeurIPS 2025) to the related work. We do not ablate against it because it requires a differentiable constraint: its Gauss-Newton projection onto equality constraints needs the residual and its Jacobian, whereas the whole point of CSFM is a setting with only a binary membership oracle over sets of positive volume. The two are complementary: PCFM corrects each sample at sampling time; CSFM learns part of the correction into the velocity field and keeps a membership-only repair for deployment.
We hope these additions strengthen the reviewer's support for the paper, and we would be glad to add any further comparisons.