- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2(−1)
- Soundness
- 2
Read review
Paper summary
This paper presents Hessian-vector product (HVP) matching to train a coarse-grained potential for molecular dynamics. The primary goal is to jointly supervise the gradient and curvature of the free energy surface, which previous force matching methods fail to capture. Through theoretical derivation, the target CG Hessian is decomposed into two terms: the first is precomputed before training, while the second is computed online at negligible cost. The in-distribution evaluations on 9 fast-folding proteins demonstrate that the Hessian correction generally improves performance in both distributional similarity and structural validity over force matching methods, with acceptable computational overhead during training.
Critical strengths
- Supervising the curvature of the free energy surface is a promising direction, and through theoretical derivation, the authors develop an effective training method with manageable computational overhead, which is critical for the practical utility of the algorithm.
- The experiments, covering both in-distribution (Section 4) and out-of-distribution settings (Appendix C), are sufficiently thorough to demonstrate the effectiveness of the proposed method.
Critical weaknesses
- The writing and organization of the paper could be improved to some extent. (i) The "Related Work" section could be presented as a separate section following the Introduction, while the contributions could be moved to the end of the Introduction for better readability. (ii) The "Discussion" section is somewhat overly long, especially given that the main text does not fully use the 9-page limit. (iii) The "Experiments" section is not well organized and is therefore difficult to follow. For example, the "Model variants" subsection includes baselines, hyperparameters, computational infrastructure, and training-cost comparisons. These could be separated into different paragraphs, with some details moved to the Appendix for better coherence. Similarly, the "Evaluation" subsection mixes evaluation metrics with experimental results, which could also be presented separately. In addition, the out-of-distribution evaluation in Appendix C is important and would be better placed in the main text.
Questions for authors
- As noted in the Weakness section, my major concern is that the writing and organization of the paper have not yet reached the standard expected for ICLR. If the authors carefully revise the manuscript, I would be happy to reconsider and potentially raise my score during the rebuttal stage.
- I have some concerns about the correspondence between Eq. (9) and Eq. (10). The paper states that Eq. (10) is obtained by multiplying both sides of Eq. (9) by . However, (i) the first term in Eq. (9) is an ensemble-averaged projected Hessian, whereas the ensemble averaging appears to be omitted in the first term of Eq. (10). (ii) The second term in Eq. (9) depends only on and is independent of the CG model, while Eq. (10) introduces a term involving . Could the authors please clarify these points?
- Equation (16) suggests that HVP matching requires at least one additional autograd operation compared with force matching, which is roughly equivalent to an extra backpropagation step. However, Lines 332–333 report only about a 16% increase in training time. Could the authors explain the reason for this relatively small overhead?
- In Table 1, for several systems such as Chignolin, Protein B, and Homeodomain, using only the AA Hessian projection appears to yield better performance in the TICA space than using the full loss. Could the authors explain why this occurs?
3 comments
Response to Reviewer Vx2H (2/2)
Q3. Why is the training overhead only about 16%?
Upon checking, we realized that we did not rigorously measure and record the training time consistently across the three variants, so the 16% figure is not properly supported. We have therefore removed these timing claims from the revised paper.
Instead, the revision describes the cost by counting operations, which does not depend on timing measurements (Section 5, Training cost). Each force-matching step already needs one backward pass to get the forces from the energy and a second backward pass to get the parameter gradients. HVP matching adds further backward passes (one per probe, here), and the parameter-gradient pass then also runs through them. So HVP matching costs more per step than force matching, and this extra cost grows with . The covariance correction (Term 2) only needs a few vector operations on forces that are already computed. Term 1 needs all-atom force evaluations per frame, done once before training and reused. We now state these costs plainly and no longer describe the overhead as small.
Q4. Why does Term 1 alone sometimes beat the full loss in TICA space (chignolin, Protein B, Homeodomain)?
We see three reasons.
- Some of these gaps are within the error bars. For Protein B (TIC0: vs. ) and Homeodomain (TIC0: vs. ; TIC1: vs. ), the two variants are not clearly different. Chignolin is a clear case (TIC0: vs. ).
- Term 2 is much noisier. It is estimated from a single frame, and the Hessian part of the training loss is roughly 75 times larger with the covariance term than without it (Fig. 2, Appendix B). This noise, plus the bias early in training, can outweigh the benefit of the more correct target in some systems.
- Term 1 alone is a too-stiff target. On the model with a known exact answer (Appendix D), Term 1 alone is up to 47% too stiff in the softest mode. A too-stiff target is not correct physics, but it can act as a stabilizing regularizer. That can improve the simulated distribution for some systems even though the learned Hessian is less accurate.
In the revision, we no longer present FM+ as an approximation of the CG Hessian. We present it as an ablation (Section 5), explain these points (Section 6), and note that each variant was trained with a single training seed, so the error bars reflect MD sampling only (Sections 5 and 7).
Thanks for the detailed rebuttal and the revisions to the manuscript. However, several of my concerns remain insufficiently addressed.
- Writing and organization. I have reviewed the revised manuscript and acknowledge the improvements to the experimental setup section. Nevertheless, I believe the overall presentation still requires substantial improvement. For example, (i) the related work could be organized into thematic subsections rather than presented as a largely continuous discussion, (ii) the main text contains only one figure and one table, making several lengthy textual explanations difficult to follow intuitively, and (iii) the conclusion occupies an entire page, whereas the experimental section, which is arguably more important for evaluating the proposed method, spans only approximately two pages. I encourage the authors to learn from the organization and presentation of previously accepted conference papers.
- Equations (9) and (10) in the original submission. This concern has been adequately addressed.
- Training overhead. The authors acknowledge in their rebuttal that the original training time measurements were not recorded rigorously. It raises broader concerns about the reliability and reproducibility of the experimental evaluation, particularly whether the training procedures and computational budgets were consistently and fairly controlled across methods.
- Effectiveness of the proposed method. Regarding Table 1, the authors explain that Term 2 produces excessively noisy supervision, whereas Term 1 provides an overly stiff training target. However, these explanations raise further questions about the effectiveness and necessity of the proposed formulation. The additional training complexity introduced by both terms does not appear to be justified by consistent performance improvements in Table 1. Therefore, I remain unconvinced that the increased optimization difficulty is warranted.
Overall, I will maintain my original score.
Response to Reviewer Vx2H (1/2)
We thank the reviewer for the clear and constructive suggestions, and for the willingness to reconsider the score. We have made the requested changes in the revised PDF, with one exception (the out-of-distribution study), which we explain below.
Writing and organization (Weakness, Q1).
(i) Related Work is now its own section (Section 2), right after the Introduction, and the contributions are listed at the end of the Introduction.
(ii) The Discussion is shorter. It is now a brief Conclusion with a short Limitations paragraph (Section 7).
(iii) The experimental setup (Section 5) is split into clear parts: Data and benchmark; Models and baselines; Training details; Training cost; Evaluation metrics. Results are in a separate section (Section 6).
Q2. Correspondence between Eq. (9) and Eq. (10).
Thank you for this comment; both observations are correct. (i) Eq. (10) of the submission is a single-frame quantity, not the ensemble average. Exactly as in force matching, the averaging is done by the squared loss over many frames. Fitting noisy targets with a squared loss returns their average. (ii) The covariance in Eq. (9) indeed depends only on all-atom forces. But computing a covariance requires the mean force, which is unknown, so we use the network force FNN (which force matching trains to approximate the mean force FCG) as the estimate. This makes the practical target depend on the model and introduces a bias of exactly βεεT, where ε is the model's force error. The bias vanishes when force matching converges. In the revision, the estimator H~CG is defined explicitly (Eq. 12) and its average is stated (Section 4.1); the analysis and a new test on a model with a known exact Hessian are in Appendix D. We explain this step in more detail in our response to Reviewer pzQP (Q3), and give the full bias analysis and the test in our response to Reviewer nFXT (Part 1).