acceptodds
Under review as a conference paper at ICLR 2027

Hessian-Informed Flow Matching

11 views

Abstract

Data from physical systems often concentrate near stable states corresponding to minima of an energy function. Around these states, the curvature of the energy landscape shapes the local covariance structure of the data, which is often highly anisotropic: small perturbations around a data point raise its energy more sharply in some directions than in others. However, standard flow matching commonly connects noise to target data through isotropic conditional paths that do not explicitly follow this physical structure. We introduce Hessian-Informed Flow Matching (HI-FM), which uses the energy Hessian at each data sample to encode this local curvature in the conditional probability paths used for training. By locally linearizing energy-driven dynamics, we derive anisotropic paths that move samples toward their targets at different rates in different directions, without changing the standard flow matching objective. We evaluate HI-FM through molecular generation on Hessian-QM9 and GEOM-Drugs, using SemlaFlow as a shared backbone. With Hessians calculated offline only for the training data, across both datasets, HI-FM improves the physical and geometric quality of generated molecules while maintaining comparable chemical validity. Under a matched optimizer-update budget, HI-FM trained with only % of Hessian-QM9 achieves lower mean energy, strain, and bond-length than the full-data SemlaFlow baseline. The results support local energy curvature as a useful inductive bias for flow matching.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
flow matching, generative modeling, molecular generation, energy Hessians

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_GisoOpenReview ↗
Rating
4(+2)
Confidence
4
Informativeness
3(+1)
Presentation
3
Soundness
3(+1)
Read review
Paper summary

This paper introduces Hessian-Informed Flow Matching (HI-FM), which incorporates the anisotropy of physical energy landscapes into flow matching. Standard linear conditional probability paths constract equally in all directions. Hi-FM instead uses energy Hessians precomputed at training samples to construct paths with direction-dependent contraction rates. This preserves the source distribution and target endpoints while allowing faster contraction along stiff, high-curvature direction. Hessians are used to construct the training paths and regression targets but are not required during generation. Experiments on Hessian-QM9 and GEOM-Drugs, using SemlaFlow as the primary backbone, show improvements in energy, strain, and bond-length distribution metrics over the baseline under matched optimizer-update budgets, while maintaining broadly comparable chemical validity. Improved alignment between the Jacobians of the learned vector fields and the energy Hessians further indicates that the model better captures local energy curvature.

Critical strengths
  • The paper addresses the lack of explicit physical anisotropy in standard linear flow matching paths.
  • The main idea is straightforward, and the paper clearly explains what the proposed change is meant to achieve.
Critical weaknesses
  • The empirical evidence is limited in scope. The curvature analysis does not cover generated molecules, and the effect of Hessian quality and its cost are not isolated. All comparisons disable OT coupling within a single backbone family, and the GEOM-Drugs results rely on Hessians computed away from MMFF94 minima and on a single training seed. As a result, it is not yet clear how robust the gains are and whether they generalize beyond the specific settings tested.
Questions for authors

My concerns are mainly about the scope of the empirical evidence rather than the core idea. I would be open to raising my score if the authors can address these questions during the rebuttal, even with preliminary results (e.g., a small-scale run or a single additional seed).

  • Does the improved Hessian-Jacobian alignment also hold for generated molecules? How do their Hessian spectra compare with those of the data?
  • Could the authors compare DFT Hessians, MMFF94 Hessians on the same Hessian-QM9 molecules and report preprocessing and total training costs?
  • Do the gains persist with OT coupling enabled in SemlaFlow? Whould HI-FM also improve other models, such as FlowMol?
  • How close are the GEOM-Drugs structures to MMFF94 minima, as measured by their gradients? Are the gains consistent across additional training seed?
3 comments

We thank the reviewer for the careful assessment and for stating which results would strengthen the paper. We address each point below.

Response to the Critical Weakness and Question 1:

To test whether Hessian-Jacobian alignment extends to generated molecules, we evaluated the 10% Hessian-QM9 models on a shared set of 1,000 generated molecules: 500 from each method. Both learned fields were evaluated at the final coordinates of every molecule in this set. We computed the diagnostic MMFF94 Hessians at these coordinates without relaxing the geometries.

ModelProjSim@10 ↑\uparrowHessianFit ↑\uparrow
SemlaFlow0.239±0.0490.239\pm0.0490.668±0.0390.668\pm0.039
HI-FM-D0.767±0.062\mathbf{0.767\pm0.062}0.745±0.041\mathbf{0.745\pm0.041}

HI-FM-D achieves substantially higher alignment on this shared set of generated geometries. Since both fields are evaluated at identical molecular coordinates, this advantage is not due to evaluating each method on a different set of generated structures.

We also checked whether the generated structures themselves have curvature statistics closer to the reference conformations. For this, we compared their normalized internal Hessian spectra, negative curvature and gradients:

SetnnInternal W1↓W_1\downarrowNegative curvature (%) ↓\downarrowGradient RMS (median) ↓\downarrow
SemlaFlow9,1150.014237.043.14
HI-FM-D9,1410.006122.933.57
Reference50001.418.87

Spectra are normalized by each molecule's largest absolute internal eigenvalue. Negative curvature reports the percentage of conformations with an internal eigenvalue below −10−2-10^{-2} on this scale. Gradient RMS is over atomic gradient magnitudes, including hydrogens, in kcal/mol/Å.

Compared with SemlaFlow, HI-FM-D reduces spectral distance by 57%, median gradient RMS by 22%, and the fraction with substantial negative curvature from 37.0% to 22.9%. Thus, the higher alignment is accompanied by generated structures with normalized curvature statistics closer to the reference. The Hessian evaluations are used for the diagnostics above; sampling itself requires no Hessian evaluations.

Response to Question 2:

Following the reviewer's suggestion, we compared HI-FM-D using DFT and MMFF94 Hessians at the same 3,100 conformations from 10% Hessian-QM9. We also tested whether the gains persist when using only the ten stiffest DFT modes or adding 25% internal noise to the DFT Hessians.

SettingEnergy/atom ↓\downarrowStrain/atom ↓\downarrowLength W1↓W_1\downarrowAngle W1↓W_1\downarrowValidity (%) ↑\uparrow
SemlaFlow2.4501.3841.8601.28994.31
HI-FM-D, DFT2.1801.0711.2601.10194.75
HI-FM-D, MMFF942.1441.0591.3101.01794.16
HI-FM-D, 10 modes2.2441.1331.5331.16994.56
HI-FM-D, 25% noise2.1551.0531.2931.05195.00

MMFF94 gives similar energy, strain and geometry gains to DFT in this comparison. Using only ten modes still retains 76% of the energy gain and 80% of the strain gain obtained with full DFT Hessians over SemlaFlow. With 25% internal noise, at least 94% of the full-DFT gain remains in each of the four energy and geometry metrics. These results show that the gains do not require a complete, unperturbed DFT Hessian.

The gains are also not limited to MMFF94 evaluation. On GEOM-Drugs, HI-FM-D trained with MMFF94 Hessians reduces mean GFN2-xTB strain by about 26% with 10% training data and 23% with full data, relative to SemlaFlow.

For preprocessing, we reuse the public DFT Hessians for the 13,607-molecule Hessian-QM9 pack without repeating quantum-chemical calculations. Their eigendecomposition takes about 7 s on one CPU core. Constructing MMFF94 Hessians at the same coordinates takes about 0.6 min, followed by eigendecomposition. The cached eigensystems occupy 0.16 GiB.

For training, we measured complete 33,000-update runs on 10% Hessian-QM9 using one H200 GPU, including periodic validation. SemlaFlow took 7.63 h, while HI-FM-D took 7.86 h with either Hessian source, a 3.0% increase. With equivariant OT, SemlaFlow took 7.82 h and HI-FM-D took 8.11 h, a 3.7% increase.

In separate A100 benchmarks with three paired seeds, the per-update overhead falls from 16.8±4.116.8\pm4.1\\% on 10% Hessian-QM9 to 2.6±1.02.6\pm1.0\\% on the full dataset. Thus, the per-update overhead is substantially smaller on full QM9 in these benchmarks.

We will add these results and protocols to the revised manuscript.

Response to Question 3:

We tested whether HI-FM's gains remain when SemlaFlow's native equivariant OT coupling is enabled. Both methods were trained on 10% Hessian-QM9 with matched training settings:

MetricSemlaFlow + OTHI-FM-D + OT
Energy/atom ↓\downarrow2.2932.103
Strain/atom ↓\downarrow1.0560.954
Length W1W_1 (10−310^{-3} Å) ↓\downarrow1.8071.460
Angle W1W_1 (∘^\circ) ↓\downarrow1.3521.069
Validity (%) ↑\uparrow93.6693.78

Compared with SemlaFlow using the same OT coupling, HI-FM-D reduces energy by 8.3%, strain by about 10%, bond-length W1W_1 by 19%, and bond-angle W1W_1 by 21%, with similar validity. The strain gain is smaller than the 22.6% with independent coupling, but HI-FM still improves the OT-coupled baseline. Its benefit is therefore not restricted to independent coupling.

Additional architectures: In the paper, we also tested an EDM-style EGNN backbone with three seeds, four data fractions, and all three HI-FM variants, using the same c=2,κ=1c=2,\kappa=1 default (Section C.1, Table 6). Across these comparisons, seed-averaged median energy decreases by 4.9-12.7% and strain by 10.1-18.2%, with higher mean validity. These results show that the gains also hold on a second backbone with the same default parameters.

All EGNN runs use independent coupling. In those experiments, “OT-FM” refers to the straight-line conditional-OT path, not explicit source-target OT coupling. Since HI-FM modifies the coordinate path, it could in principle also be adapted to models such as FlowMol while retaining their categorical training objectives. We have not tested FlowMol, so our evidence beyond SemlaFlow comes from the EGNN results.

Response to Question 4:

GEOM-Drugs conformations were optimized with GFN2-xTB, so nonzero MMFF94 gradients are expected. Their mean MMFF94 energy and strain are about 1.1 and 0.36 kcal/mol/atom (Appendix B.6). The low strain indicates that they are close to MMFF94 minima in energy, but does not imply that their MMFF94 gradients are zero. Following the reviewer's suggestion, we computed MMFF94 gradients on 2,000 GEOM-Drugs conformations. The median RMS atomic gradient magnitude, including hydrogens, is 16.53 kcal/mol/Å, compared with 18.25 for DFT-optimized Hessian-QM9.

HI-FM does not require the data to be MMFF94 minima. Its positive contraction rates keep the conditional path endpoint at the observed conformation, even when ∇E(x1)≠0\nabla E(x_1)\ne0. A full linearization of the MMFF94 force would also include the constant term −∇E(x1)-\nabla E(x_1). We instead use the local Hessian to modify the flow-matching paths towards the observed conformation, not to reproduce MMFF94 relaxation.

As requested, we also repeated the 10% GEOM-Drugs experiment with a new training seed to check whether the gains hold. Both methods used matched training and evaluation settings:

MetricOriginal experiment (SemlaFlow)Original experiment (HI-FM-D)New training seed (SemlaFlow)New training seed (HI-FM-D)
Energy/atom ↓\downarrow2.0601.7832.0141.694
Strain/atom ↓\downarrow1.3181.0811.2881.010
GFN2-xTB strain/atom ↓\downarrow0.9390.6920.8680.661
Validity (%) ↑\uparrow90.4791.4992.2992.14
Mol. stability (%) ↑\uparrow95.7194.9996.7696.36
Length W1W_1 ↓\downarrow0.8130.6030.7470.597
Angle W1W_1 ↓\downarrow1.6401.6381.4271.494

The energy, strain and bond-length gains hold with the new seed: HI-FM-D reduces these metrics by about 16%, 22% and 20%, respectively. Mean GFN2-xTB strain is also about 24% lower, so the strain improvement holds under a different evaluation potential. Bond-angle W1W_1 is nearly unchanged in the original experiment but slightly worse with the new seed. Validity and molecular stability are also comparable in the new run.

Thank you for addressing my questions. I view the paper’s main idea positively, and my only concern was the limited experimental evidence. The additional results provide some reassurance regarding robustness across random seeds and the fairness of the experimental settings. On this basis, I support acceptance.

Reviewer_MeymOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces Hessian-Informed Flow Matching (HI-FM), which incorporates local energy-landscape information into flow matching by adapting conditional probability paths using the Hessian of the energy function at each training target. The resulting paths contract more rapidly along high-curvature directions than along low-curvature ones, providing a physics-informed inductive bias while largely preserving the standard flow-matching architecture, objective, and inference procedure. Since the Hessians are precomputed and used only during training, no additional energy or Hessian evaluations are required at inference. Experiments with SemlaFlow on Hessian-QM9 and GEOM-Drugs show consistent improvements in energy, strain, and local geometric quality under matched optimization budgets.

Critical strengths
  1. The methodological contribution is clear and well motivated. Rather than introducing local energy curvature through an additional loss term, HI-FM directly incorporates Hessian information into the design of conditional probability paths, providing a conceptually appealing way to inject physical information into flow matching. The construction preserves the source and target endpoints while establishing a direct connection between local energy-landscape geometry and flow-matching supervision.

  2. The experiments provide meaningful evidence for the proposed mechanism. In particular, the eigendirection-randomization ablation shows that preserving the Hessian spectrum while randomizing the active eigendirections removes approximately 76% of the energy improvement and 72% of the strain improvement over SemlaFlow, supporting the importance of target-specific curvature directions. Additional experiments with an EGNN backbone further suggest that the benefits of HI-FM are not specific to SemlaFlow.

Critical weaknesses
  1. The reliance on target-specific Hessians introduces non-trivial preprocessing and computational overhead. Computing and storing Hessians and their eigendecompositions may become costly for larger molecular systems. Even with cached eigensystems, the reported implementation increases the per-update training time by 18.9%, excluding the additional cost of Hessian computation itself. This requirement may limit the scalability and practical applicability of HI-FM to larger or more complex molecular systems.

  2. The physical interpretation of the GEOM-Drugs results is limited by the fidelity of the Hessian information. The Hessians for GEOM-Drugs are derived from the MMFF94 force field rather than higher-fidelity quantum-mechanical calculations. Therefore, the reported improvements demonstrate the benefit of incorporating local curvature information from the chosen force field, but do not necessarily establish that HI-FM captures the underlying PES curvature of chemically diverse molecules. Evaluating the method with higher-fidelity or alternative sources of curvature information would provide stronger evidence for its physical interpretation and generality.

Questions for authors
  1. Could the authors clarify the total offline computational and storage cost associated with Hessian computation and eigendecomposition relative to standard SemlaFlow training? In particular, it would be helpful to report the overall preprocessing overhead in addition to the per-update training cost after the eigensystems have been cached.

  2. How sensitive are the improvements of HI-FM to the quality and completeness of the Hessian information? For example, would lower-cost approximate Hessians, partial eigenspectra, or noisy Hessian estimates retain a substantial fraction of the reported gains? Clarifying this would help assess the scalability and practical applicability of HI-FM to larger molecular systems.

2 comments

We thank the reviewer for their time and helpful comments.

Response to Weakness 1 and Question 1:

We agree that preprocessing and storage costs should be reported alongside training time. For both datasets, Hessians and their eigensystems are precomputed offline, and training uses the cached eigensystems.

GEOM-Drugs contains 5,651,202 geometries and occupies about 472 GB (440 GiB), including 455.8 GB (424.5 GiB) of Hessians. The training split contains 4,522,204 geometries and occupies about 378 GB. We will release the dataset with its Hessians upon acceptance.

On one AMD EPYC 7702 CPU core, MMFF94 Hessian construction takes 0.71±0.180.71\pm0.18 s per conformation, or 0.770.77 s including RDKit molecule construction and MMFF parameter setup. Including data loading, we estimate about 1,700 CPU-core-hours for all 5.65M conformations, or about 1,400 for the training split. Conformations can be processed independently, so under ideal scaling the full dataset would take about 17 hours on 100 cores or 1.7 hours on 1,000 cores.

This preparation is done once and reused across training runs, rather than repeated for each model. For context, training one GEOM-Drugs model uses about 200 GPU-hours on A100 GPUs, including validation (Appendix B.2). Generation requires no energy or Hessian evaluations.

For Hessian-QM9, we reuse the DFT Hessians supplied with the public dataset, so no quantum-chemical calculations are repeated. For the 13,607-molecule pack, eigendecomposition takes about 7 s on one CPU core. Constructing MMFF94 Hessians at the same coordinates takes about 0.6 min, followed by eigendecomposition. The cached eigensystems occupy 0.16 GiB.

To measure the additional cost during training, we timed GEOM-Drugs updates, including data loading and path construction. Across eight matched timing windows covering four data fractions and two training stages on four GPUs, we found no consistent increase for HI-FM-D. An initial profile of the additional training seed gives 0.1769 s per update for SemlaFlow and 0.1807 s for HI-FM-D.

This small measured overhead on GEOM-Drugs differs from the 18.9% increase reported for 10% Hessian-QM9 in Table 9. To investigate this difference, we repeated the 10% and full-QM9 benchmarks with three paired seeds, keeping the batch size and four data-loading workers fixed. Both methods in each pair ran on the same A100 GPU. The table reports the change in update time for HI-FM-D relative to SemlaFlow:

DatasetBenchmarkUpdate-time change (%)
GEOM-Drugs, 10–100%Mean of eight windows−1.8-1.8
GEOM-Drugs, 10%Additional seed, initial profile+2.1+2.1
Hessian-QM9, 10%Three paired seeds+16.8±4.1+16.8 \pm 4.1
Hessian-QM9, 100%Three paired seeds+2.6±1.0+2.6 \pm 1.0

The repeated 10% QM9 benchmark gives an overhead close to the original 18.9%, but this falls to 2.6% on full QM9. In absolute terms, the extra time drops from 52 ms to 7 ms per update. Thus, the larger overhead reported in Table 9 does not persist in the full-dataset benchmark.

Less frequent data-loader restarts may explain this drop. In these runs, workers restart every 11 updates on 10% QM9, but every 141 updates on full QM9. Longer epochs spread the cost of starting workers and preparing the first batch over more updates. The 10% GEOM-Drugs split has 5,228 updates per epoch, so restarts are much less frequent there. This is consistent with its small measured overhead.

We also measured complete training-job times for the matched 10% Hessian-QM9 comparison. On one H200 GPU, 33,000-update runs took 7.63 h for SemlaFlow and 7.86 h for HI-FM-D with either Hessian source, a 3.0% increase. These timings include periodic validation but exclude the one-time preprocessing described above.

We will include these cost measurements and the full preprocessing and timing protocols in the revised manuscript.

Response to Weakness 2 and Question 2:

Following the reviewer's suggestion, we tested whether the gains persist with approximate or incomplete Hessian information. We compared full DFT Hessians with MMFF94 Hessians, a 10-mode DFT approximation, and DFT Hessians with 25% internal noise. All runs use the same 3,100 conformations from 10% Hessian-QM9, with unchanged coordinates and matched training and evaluation settings.

The 10-mode version keeps the ten stiffest DFT directions and rescales their rates to [1,2][1,2] using the retained spectrum. Other directions follow the linear path. For the noise test, we use a fixed internal Gaussian perturbation per molecule with ∣E∣F=0.25∣H∣F\\|E\\|_F=0.25\\|H\\|_F.

SettingEnergy per atom ↓\downarrowStrain per atom ↓\downarrowLength W1↓W_1 \downarrowAngle W1↓W_1 \downarrowValidity (%) ↑\uparrow
SemlaFlow2.4501.3841.8601.28994.31
HI-FM-D, DFT2.1801.0711.2601.10194.75
HI-FM-D, MMFF942.1441.0591.3101.01794.16
HI-FM-D, 10 modes2.2441.1331.5331.16994.56
HI-FM-D, 25% noise2.1551.0531.2931.05195.00

MMFF94 gives similar energy, strain and geometry gains to full DFT Hessians in this matched comparison. Its Hessians take about 3 ms per QM9 molecule on one CPU core in our implementation.

Using only ten modes still keeps about 76% of the energy gain and 80% of the strain gain obtained with full DFT Hessians over SemlaFlow. The bond-length and bond-angle metrics also improve, but less than with full DFT Hessians. We select these modes from the full DFT eigensystem, so this test assesses the effect of using fewer modes; it does not reduce Hessian construction cost.

The gains also persist in the noise test. With 25% internal noise, at least 94% of the gain obtained with full DFT Hessians over SemlaFlow remains in each of the four energy and geometry metrics.

Since the QM9 energy and strain metrics themselves use MMFF94, we also checked whether the gains extend to another evaluation potential. We evaluated generated GEOM-Drugs molecules with GFN2-xTB. HI-FM-D trained with MMFF94 Hessians reduces mean GFN2-xTB strain by about 26% and 24% for two 10%-training seeds, and 23% with full training data, relative to SemlaFlow. Thus, the strain improvement is not limited to MMFF94 evaluation.

The Hessian ablations show that these gains do not require a complete, unperturbed DFT Hessian. Of course this does not mean that the directions are unimportant: the randomization test in Table 10 shows that target-specific directions still matter.

We will add these ablations and their protocols to the revised manuscript.

Reviewer_qhVyOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a method to incorporate the local curvature of a physical energy landscape into the conditional probability paths of flow matching. This method named Hessian-Informed Flow Matching (HI-FM) was demonstrated to achieve better generation quality, data efficiency, and optimisation efficiency for the 3D molecular generation task.

The paper argues, that in contrast to the standard linear conditional paths (in flow matching) that contract noise isotropically at the same rate in all directions, this method uses precomputed energy Hessians to produce anisotropic paths that contract high-curvature ("stiff") directions faster than low-curvature ("soft") directions. As HI-FM only alters the conditional paths and their velocity targets, the neural network architecture, training objective, and inference-time sampling procedure remain unchanged, requiring no energy or Hessian evaluations during generation.

Although the evaluation in the paper is restricted to molecular generation tasks, this procedure can be applied to any training procedure employing flow matching for noise estimation.

Critical strengths
  1. HI-FM incorporates local energy curvature into the generative process simply by modifying the conditional training paths and velocity targets, requiring no changes to the underlying neural network architecture or the standard flow matching objective. Because the target Hessians are precomputed offline exclusively for the training set, generating new molecules at test time requires no energy or Hessian evaluations and adds zero inference latency. Furthermore, the randomized-eigendirection ablation (Table 10) shows that rotating the active Hessian eigendirections while preserving the spectrum removes approximately 76% of the energy gain and 72% of the strain gain. This indicates that the specific curvature directions, not just the spectrum, drive the improvement. The Jacobian–Hessian alignment (Table 3) further confirms that the network absorbs this curvature supervision. This combination of a simple, drop-in change with evidence that the gains come from the proposed mechanism is the main reason I support acceptance.

  2. Under the same optimizer-update budget (33,000 steps), SemlaFlow trained with HI-FM on only 10% of Hessian-QM9 achieves lower mean energy (2.098 vs. 2.203 kcal/mol/atom) and strain (0.999 vs. 1.083 kcal/mol/atom) than the full-data (100%) SemlaFlow baseline (Table 1). Within the 10% setting, it also reaches the best validation energy and strain of SemlaFlow trained on the same 10% subset with 23–31% fewer optimizer updates (Figure 2e,f). This demonstrates that local curvature information can effectively compensate for limited training conformations, which is practically significant for settings where high-quality conformational data are scarce.

Critical weaknesses
  1. Discrepancy Between Theoretical Motivation and Heuristic Rescaling (Hyperparameter Sensitivity): The method is motivated by exact linearized gradient relaxation rates λi\lambda_i (Section 4.1). However, the actual implementation linearly compresses all active positive eigenvalues into a narrow interval bounded by c=2c=2, so directional exponents lie in ri∈[1,2]r_i \in [1, 2] (Appendix B.4). The sensitivity ablation in Table 11 (Appendix D.3) reveals two problems. Allowing a wider spread of relaxation rates (c=5c=5) raises energy above the SemlaFlow baseline (2.7052.705 vs. 2.3482.348 kcal/mol/atom). Changing the time-scaling parameter in either direction (κ=0.5\kappa=0.5 or κ=2\kappa=2) also degrades performance, severely so at κ=2\kappa=2 (5.9995.999 kcal/mol/atom). This sensitivity outside a narrow parameter window around (c=2,κ=1)(c=2, \kappa=1) suggests that the method helps mainly as a mild anisotropic perturbation of the linear path. It weakens the claim that the paths naturally capture multi-timescale physical dynamics.

  2. Experimental Caveats (Modified Baseline, Metric Overlap, and Unreported Preprocessing Costs):

a. Modified baseline: To isolate the coordinate path construction, the authors disabled SemlaFlow's built-in coordinate Optimal Transport (OT) coupling and coordinate-path noise across all runs (Appendix B.2), so the SemlaFlow baseline is weaker than its published configuration. The EGNN experiment (Section C.1) does compare against an OT-FM baseline. However, it is unclear whether the HI-FM variants there also use OT coupling, so it remains open whether HI-FM's gains persist when both methods use OT-coupled training.

b. Metric overlap: For the GEOM-Drugs experiments, which evaluate a single training seed (Table 2), the MMFF94 force field is used both to compute the training Hessians and to evaluate the test energy and strain of generated molecules (Sections B.1.2 and B.6). Using the same empirical force field for both supervision and evaluation creates a favorable alignment that may overstate general physical accuracy. This concern is partly mitigated on Hessian-QM9, where supervision uses DFT Hessians and evaluation uses MMFF94. Still, the larger-molecule results rest on a single seed with matched force fields.

c. Unreported costs: Per-update training overhead with cached eigensystems is reported only for 10% of Hessian-QM9 (+18.9%, Table 9). The offline computation time and disk storage required for the 3N×3N3N \times 3N Hessian eigensystems on GEOM-Drugs are not quantified.

Questions for authors

(Clarification) Performance degradation at c=5c=5 and κ≠1\kappa \neq 1 (Table 11): Could the authors clarify why increasing cc to 55, or changing κ\kappa to 0.50.5 or 22, causes energy and strain to degrade below the standard SemlaFlow baseline? For κ=2\kappa=2, all exponents satisfy ri<1r_i < 1, so the velocity target ri(1−t)ri−1r_i(1-t)^{r_i-1} diverges as t→1t \to 1 and most of the displacement is deferred to the end of the path. Does the degradation stem from discretization errors in the fixed 100-step Euler sampler, from the interaction with Beta(2,1) time sampling and the endpoint-weighted loss, or from optimization instability caused by larger target velocities during training?

(Clarification) Comparison with Optimal Transport (OT) coupling: Since SemlaFlow's coordinate OT coupling was disabled for the main experiments (Appendix B.2), how does standard SemlaFlow with its native OT coupling enabled compare against HI-FM on Hessian-QM9? In the EGNN experiments (Section C.1), do the HI-FM variants also use OT coupling, or only the OT-FM baseline?

(Suggestion) Compatibility with OT alignment: Since the HI-FM map (Eq. 7) accepts any source sample x0x_0 and PP depends only on x1x_1, rigid-body or rotational OT alignment of x0x_0 appears compatible with the anisotropic path in principle. Have the authors tried combining the two?

(Clarification) Preprocessing, storage, and overhead for GEOM-Drugs: What is the total wall-clock time and disk storage footprint required to precompute and store the MMFF94 Hessians and eigendecompositions for the full GEOM-Drugs training split? Section D.1 mentions caching eigendecompositions on Hessian-QM9. Were they also cached for GEOM-Drugs, and what is the per-update training overhead there?

3 comments

We thank the reviewer for their time and valuable comments. We address each point with additional experiments and clarifications below.

Response to Weakness 1 and Question 1:

The rescaling changes the physical relaxation rates, but preserves the Hessian eigendirections and their stiffness ordering. HI-FM uses this information to modify the flow-matching paths, but does not aim to reproduce molecular trajectories. So we do not need to preserve the raw eigenvalue ratios. In fact, the active Hessian eigenvalues span several orders of magnitude (median ratio 3.1×1033.1\times10^3 on Hessian-QM9). Using their raw ratios would give extreme contraction rates that would be numerically difficult to learn and integrate. We therefore rescale the rates, with cc and κ\kappa to control the range and timing of contraction conducive to better learning as opposed to being faithful to the true dynamics.

We believe our default values c=2,κ=1c=2,\kappa=1 balances the physical information with controlled contraction. As the reviewer has mentioned, the rates stay in [1,2][1,2], which keeps conditional velocities bounded and have the softest modes follow the linear schedule, and the stiffer modes contract earlier. It should be noted that the same setting improves energy and strain with both SemlaFlow and EGNN backbones across datasets, suggesting potential universality at least for our problem setting.

Following the reviewer's suggestion, we reevaluated the same 10% Hessian-QM9 checkpoints with 100 and 500 Euler steps to test whether numerical integration contributes to the exhibited sensitivity:

SettingEnergy/atom ↓\downarrow (100 steps)Energy/atom ↓\downarrow (500 steps)Strain/atom ↓\downarrow (100 steps)Strain/atom ↓\downarrow (500 steps)
SemlaFlow2.3382.3001.2731.301
HI-FM-D, c=2,κ=1c=2,\kappa=12.1331.9811.0310.967
HI-FM-D, c=1.5,κ=1c=1.5,\kappa=12.2492.1041.1271.059
HI-FM-D, c=5,κ=1c=5,\kappa=12.7872.5961.7161.585
HI-FM-D, c=2,κ=0.5c=2,\kappa=0.52.5832.1791.3101.133
HI-FM-D, c=2,κ=2c=2,\kappa=26.2475.9435.1414.810

With κ=0.5\kappa=0.5, contraction happens early, where our Euler grid is coarser. At 500 steps, it outperforms SemlaFlow in both metrics, suggesting that numerical integration partly explains its poor result at 100 steps. The c=5c=5 and κ=2\kappa=2 settings also improve but remain worse than SemlaFlow. With c=5c=5, the stiffest modes contract five times faster than the softest, which makes the regression targets harder; finer steps recover only part of the gap.

For κ=2\kappa=2, contraction is delayed towards t=1t=1, so we also tested a later endpoint cutoff at 100 steps. Moving the cutoff from 0.99 to 0.9999 reduces strain from 5.141 to 3.170, closing nearly half of the gap to the default parameter. We noticed some conditional velocities diverge near t=1t=1, but the coordinate targets used in training remain finite.

Importantly, the existing ablation in Table 10 already tests whether the directions matter. Randomizing them while keeping the eigenvalues removes 76% of the energy gain and 72% of the strain gain. So the benefit depends on target-specific directions, not just on using different rates.

To reiterate, despite the sensitivity of the quantitative results, c=2,κ=1c=2,\kappa=1 consistently improves energy and strain across the evaluated backbones and sampling budgets, supporting its use as a practical setting.

Response to Weakness 2a and Questions 2-3:

Our original comparison used independent coupling and zero coordinate-path noise to isolate the effect of changing the conditional path (Appendix B.2). As the reviewer points out, this leaves open whether the gains persist when OT coupling is enabled. Following the reviewer's suggestion, we therefore trained both SemlaFlow and HI-FM-D with SemlaFlow's native equivariant OT coupling on 10% Hessian-QM9, using matched training settings:

MethodEnergy ↓\downarrowStrain ↓\downarrowValidity (%) ↑\uparrowLength W1W_1 ↓\downarrowAngle W1W_1 ↓\downarrow
Independent coupling
SemlaFlow2.4501.38494.3131.8601.289
HI-FM-D2.1801.07194.7471.2601.101
Equivariant OT coupling
SemlaFlow2.2931.05693.6631.8071.352
HI-FM-D2.1030.95493.7831.4601.069

Enabling OT improves SemlaFlow's energy and strain. Importantly, HI-FM-D reduces them further by 8.3% and 9.7%, respectively, when both methods use the same OT coupling. The strain gain is smaller than the 22.6% with independent coupling, but these results show that HI-FM can be combined with native OT coupling and still improve the OT-coupled baseline. Its benefit is therefore not restricted to the original independent-coupling setting.

To clarify the EGNN experiments, both the “OT-FM” baseline and all HI-FM variants use independent coupling. The label “OT-FM” refers to the straight-line conditional path (Lipman et al., 2023), not to explicit source–target OT coupling (Tong et al., 2024). We will make this distinction explicit and add the matched OT comparison above to the revised manuscript.

Response to Weakness 2b:

We agree that using MMFF94 for both training Hessians and evaluation on GEOM-Drugs leaves open whether the gains extend beyond that force field. To test this, we also measured the strain of generated molecules with GFN2-xTB, a semiempirical quantum-chemical method different from MMFF94. The GEOM-Drugs models still use MMFF94 training Hessians, so this tests whether their gains hold under a different evaluation potential. Strain is reported in kcal/mol/atom.

SettingMethodMean strain ↓\downarrowMedian strain ↓\downarrow
GEOM-Drugs 10%SemlaFlow0.9390.761
HI-FM-D0.6920.567
GEOM-Drugs 100%SemlaFlow0.9160.726
HI-FM-D0.7040.565
Hessian-QM9 10%SemlaFlow1.7510.737
HI-FM-D1.3910.491

HI-FM-D has lower mean and median GFN2-xTB strain in all three settings. On GEOM-Drugs, mean strain is about 26% lower than SemlaFlow with 10% training data and 23% lower with full training data. These results show that the gains are not limited to evaluation with the same force field used to compute the training Hessians.

To address the single-seed concern, we also repeated the 10% GEOM-Drugs experiment with a new training seed. Mean GFN2-xTB strain again decreases, from 0.868 for SemlaFlow to 0.661 kcal/mol/atom for HI-FM-D. This is a reduction of about 24%, compared with 26% for the original seed, so the strain improvement holds in both runs.

We will add these results and the evaluation protocol to the revised manuscript.

Response to Weakness 2c and Question 4:

As the reviewer points out, per-update timings alone do not capture the one-time preprocessing cost. We therefore report the preparation and storage requirements below, along with additional training-time measurements. For both datasets, Hessians and their eigensystems are precomputed offline, and training uses the cached eigensystems.

GEOM-Drugs contains 5,651,202 geometries and occupies about 472 GB (440 GiB), of which 455.8 GB (424.5 GiB) stores the Hessians. The training split contains 4,522,204 geometries and occupies about 378 GB. We will release the dataset with its Hessians upon acceptance.

On one AMD EPYC 7702 CPU core, MMFF94 Hessian construction takes 0.71±0.180.71\pm0.18 s per conformation, or 0.770.77 s including RDKit molecule construction and MMFF parameter setup. Including data loading, we estimate about 1,700 CPU-core-hours for all 5.65M conformations, or about 1,400 for the training split. Conformations can be processed independently, so under ideal scaling the full dataset would take about 17 hours on 100 cores or 1.7 hours on 1,000 cores.

This preparation is done once and reused across training runs, rather than repeated for each model. For context, training one GEOM-Drugs model uses about 200 GPU-hours on A100 GPUs, including validation (Appendix B.2). Generation requires no energy or Hessian evaluations.

To measure the additional cost during training, we followed the reviewer's suggestion and timed GEOM-Drugs updates, including data loading and path construction. Across eight matched timing windows covering four data fractions and two training stages on four GPUs, we found no consistent increase for HI-FM-D. An initial profile of the additional training seed gives 0.1769 s per update for SemlaFlow and 0.1807 s for HI-FM-D.

This small measured overhead on GEOM-Drugs differs from the 18.9% increase reported for 10% Hessian-QM9 in Table 9. To investigate this difference, we repeated the 10% and full-QM9 benchmarks with three paired seeds, keeping the batch size and four data-loading workers fixed. Both methods in each pair ran on the same A100 GPU. The table reports the change in update time for HI-FM-D relative to SemlaFlow:

DatasetBenchmarkUpdate-time change (%)
GEOM-Drugs, 10–100%Mean of eight windows−1.8-1.8
GEOM-Drugs, 10%Additional seed, initial profile+2.1+2.1
Hessian-QM9, 10%Three paired seeds+16.8±4.1+16.8 \pm 4.1
Hessian-QM9, 100%Three paired seeds+2.6±1.0+2.6 \pm 1.0

The repeated 10% QM9 benchmark gives an overhead close to the original 18.9%, but this falls to 2.6% on full QM9. In absolute terms, the extra time drops from 52 ms to 7 ms per update. Thus, the larger overhead reported in Table 9 does not persist in the full-dataset benchmark.

Less frequent data-loader restarts may explain this drop. In these runs, workers restart every 11 updates on 10% QM9, but every 141 updates on full QM9. Longer epochs spread the cost of starting workers and preparing the first batch over more updates. The 10% GEOM-Drugs split has 5,228 updates per epoch, so restarts are much less frequent there. This is consistent with its small measured overhead.

We will include these cost measurements and the full preprocessing and timing protocols in the revised manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.