acceptodds
Under review as a conference paper at ICLR 2027

Wasserstein Sampling via Bridge Interpolants

Abstract

Langevin dynamics provides a classical route to sampling from unnormalized densities by evolving a distribution along the Wasserstein gradient flow of its Kullback–Leibler divergence to the target. We introduce Wasserstein Sampling via Bridge Interpolants (WASABI), a method for amortizing this sampling process using diffusion models. WASABI translates Wasserstein descent steps in distribution space into updates of the diffusion model's drift, learned through regression on samples generated by the model itself. We derive unbiased estimators that exhibit low variance empirically, making these updates trainable without evaluating the current model's density or score. The resulting fixed-point algorithm alternates between regression updates using a frozen teacher and sample generation from the updated model, with a replay buffer reusing generated samples and their expensive energy-gradient evaluations. One alanine dipeptide, WASABI outperforms reported baselines using an estimated times fewer energy evaluations, while remaining competitive on interacting particle systems benchmarks.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Sampling, diffusion models, neural samplers, Wasserstein gradient flows, langevin dynamics

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_yR4QOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

WASABI learns a diffusion sampler for an unnormalized Boltzmann distribution by translating infinitesimal Wasserstein descent into updates of a bridge denoiser. It estimates mean and covariance contributions using generated samples and target energy gradients, alternating denoiser fitting, correction learning, and sample generation with replay. Experiments on DW-4, LJ-13, and alanine dipeptide report improved backbone-angle Jensen–Shannon divergence on the molecular benchmark.

Critical strengths

1. A substantive connection between distributional transport and denoiser updates. Section 3.1 and Appendix A.1 derive the first-order response of the bridge posterior mean to endpoint transport, accounting for both endpoint displacement and changes in conditional weights. Under the stated regularity assumptions, this provides a concrete mechanism for translating a prescribed Wasserstein direction into an update of a generative model, rather than merely invoking an analogy with Langevin dynamics.

2. A tractable, score-free regression formulation. Sections 3.2–3.3 and Appendices A.3–A.5 make the regression objective computable from generated endpoints, cached target gradients, and network evaluations without evaluating the sampler's density or score. The covariance estimator supplies the expected inner product needed for training, rather than a pointwise covariance estimate, avoiding a separate conditional-sampling procedure under the stated exact-denoiser and integrability assumptions.

3. A meaningful improvement on the reported molecular observable. Section 4.3 and Table 3 report ALA2 backbone-angle Jensen–Shannon divergence of 0.0254 ± 0.0033, compared with 0.068 ± 0.009 for damped BMS [1], across three seeds and approximately one million generated samples per seed. This is a substantial improvement on that metric, while the matched DW-4 experiment in Section 4.1 and Table 1 also gives WASABI the lowest mean energy Wasserstein distance among the compared methods.

Critical weaknesses

1. The practical loss-to-theory gap and its bias–variance consequences are not adequately validated. Section 3.1 establishes the first-order Wasserstein correspondence for full covariance weight and a common infinitesimal step, whereas the experiments in Appendix B.4 downweight covariance; moreover, the estimator's unbiasedness in Section 3.2 assumes an exact conditional mean that is replaced by a learned denoiser. Downweighting preserves the target's ideal fixed point but changes the update direction, while denoiser approximation can bias the estimated correction; neither this bias nor the tilted estimator's stochastic-gradient variance is quantified in Sections 3.2–3.3 or Appendix A.4. The issue is therefore not a demonstrated algebraic error in the loss, but an unvalidated connection between minimizing the practical objective and obtaining accurate, statistically efficient Wasserstein updates.

2. The empirical advantage is mixed, and the benefit of the covariance correction is not isolated. Section 4.2 and Table 2 show that WASABI trails BMS [1] on both LJ-13 metrics: configuration Wasserstein distance is 1.597 ± 0.003 versus 1.54 ± 0.00, and energy Wasserstein distance is 1.677 ± 0.224 versus 0.53 ± 0.04; these are quoted baseline results rather than matched retraining, and the DW-4 advantage in Table 1 is limited to the mean energy metric. The strongest molecular result in Section 4.3 evaluates only the two-dimensional backbone-angle distribution despite generating Cartesian coordinates, and Appendix B.3 documents rare-region underrepresentation, so the headline result does not establish accurate sampling of the full molecular distribution. ALA2 also combines the proposed update with a high-temperature MD prior, alignment, annealing, and the bond–angle split architecture described in Appendix B.3, while the covariance sweep in Appendix B.4 and Table 5 fixes the update scale at 1.00 rather than the selected 0.10, leaving the covariance contribution to the headline gain unresolved.

3. Larger-system scalability and end-to-end efficiency remain unvalidated. Section 4 restricts evaluation to DW-4, LJ-13, and ALA2, although BMS already evaluates LJ-55 and ALA4 in its Sections 4.2–4.3 [1]; there is consequently no larger-system evidence that the additional covariance estimation and repeated denoiser fitting remain beneficial at comparable accuracy. Multi-molecule evaluation such as GEOM-DRUGS, used by Torsional Diffusion [2], is also absent, leaving broader applicability untested, although conformer generation is a different evaluation regime and LJ-55 or ALA4 is the more direct scalability test. Even the reported small-system efficiency advantage is conditional: Appendices B.3–B.4 pair accuracy from a replica-exchange-derived prior with a 3.13 million force-call estimate for a different local-MD prior, and the FLOP accounting excludes geometric calculations and other operations. This does not establish the end-to-end accuracy–cost advantage of the procedure that actually produced the reported results.

Questions for authors

1. Fidelity and estimator accuracy. Please specify the actual time-dependent update and covariance-weight schedules and, on a tractable target with a known denoiser response, measure correction error using exact versus learned denoisers and full versus downweighted covariance. Can the authors also compare the direct and tilted estimators' stochastic-gradient variance at matched computational cost as dimension and bridge time vary? This would separate deliberate changes to the update from estimator bias, sampling noise, and fitting error.

2. Contribution of covariance and molecular accuracy. At the selected ALA2 update scale of 0.10, can the authors compare zero versus the selected nonzero covariance weight while holding the prior, alignment, bond–angle split architecture, annealing, replay settings, and computational budget fixed? Please report energy-distribution agreement and minority-basin populations alongside backbone-angle divergence, with a complementary structural metric such as the TICA-based Wasserstein distance reported by BMS [1]. This would test whether covariance drives the headline improvement and whether the gain extends beyond the angle histogram.

3. Larger-system accuracy and measured cost. Can the authors evaluate one larger fixed-target system, such as LJ-55 or ALA4, against BMS [1] using matched target and evaluation conventions, and report accuracy together with end-to-end runtime, peak memory, and force evaluations? Please include the actual prior preparation, denoiser fitting, correction training, and sample-generation costs, rather than pairing measured accuracy with a different prior-generation procedure. This would establish whether the additional estimation and refitting machinery provides a practical advantage beyond the small systems studied.


References

[1] Bridge Matching Sampler: Scalable Sampling via Generalized Fixed-Point Diffusion Matching

[2] Torsional Diffusion for Molecular Conformer Generation

5 comments
AuthorsOpenReview ↗

(1/n) Response to reviewer yR4Q

We thank the reviewer for their detailed assessment and recognition of our theoretical connection, tractable formulation, and ALA2 gains. Below, we clarify the scope of our claims and provide additional evidence addressing their concerns.

W1 Theory–practice gap

We thank the reviewer for highlighting the distinction between the idealized update and the practical algorithm. The role of Theorem 3.1 is to guide practical diffusion-sampler design through Wasserstein gradient flow in distribution space. The ideal Wasserstein update requires the unknown score of the current sampling distribution and is therefore not directly tractable. By characterizing the induced first-order denoiser change through its mean and covariance contributions, the theorem guides our construction of a tractable regression update that avoids evaluating the current distribution’s density or score.

We agree that this result does not establish convergence or monotonic KL decrease for the implemented procedure. Finite step sizes, approximate neural-network fitting, and replay-buffer reuse introduce deviations from the ideal population iteration. Covariance downweighting further modifies the ideal Wasserstein direction; although it preserves the target as an ideal fixed point, this does not establish convergence. Our experiments provide evidence of practical effectiveness on the reported sampling metrics, while convergence guarantees remain an open direction for future work.

We have revised the Conclusion to clarify this scope and the theorem’s role in guiding algorithm design and have added an expanded discussion on covariance damping and the use of a replay buffer (see response to reviewer H1H8 ).

W2/Q2: empirical gains and covariance contributions

Thank you for suggesting this! Following your suggestion, we ran an ablation testing whether the model performs better with a modest covariance weight. We present the results below and find that across 3 seeds the dihedral JS to the MD reference improves by roughly 9%.

Covariance weightDihedral JS ↓Energy W2, kBT ↓
00.0579 ± 0.00441.3907 ± 0.5527
0.25 (selected)0.0527 ± 0.00571.3173 ± 0.5861

Per your suggestion, we assess structural agreement beyond the dihedral histogram by evaluating TICA-(W_2). We use the pretrained TICA model distributed with the ALA2 dataset referenced by BMS’s public implementation, without refitting. We project 10,000 generated configurations and 10,000 FAB validation configurations onto its first two components and compute their Wasserstein-2 distance.

MethodDihedral JS ↓TICA W2 ↓
WASABI0.0224 ± 0.00330.1856 ± 0.0040
ASBS, undamped0.232 ± 0.0181.263 ± 0.140
AS——
BMS, damped (λ = 10)0.068 ± 0.0090.900 ± 0.066
BMS, undamped (λ = 0)0.311 ± 0.0502.707 ± 0.268

We hope that these additional evaluations provide a more comprehensive assessment of WASABI's sampling accuracy beyond the backbone dihedral-angle distribution. In particular, the TICA-(W_2) results suggest that WASABI's improvements extend to agreement along slow collective coordinates, providing complementary evidence of its ability to capture the molecular conformational distribution. We acknowledge that these metrics do not fully characterize the high-dimensional target distribution, particularly the coverage of rare conformational states, and will clarify these limitations in the revised manuscript. We thank the reviewer for encouraging us to strengthen our empirical evaluation.

AuthorsOpenReview ↗

(2/n) Response to reviewer yR4Q

W2 Empirical Advantage

Thank you for these comments and for the careful review of the experiments. We agree that BMS achieves better LJ-13 results. However, WASABI remains competitive with strong sampling baselines: its configuration Wasserstein distance is close to ASBS (1.597 versus 1.59), and its reported mean energy Wasserstein distance is lower than both ASBS and AS (1.677 versus 1.99 and 2.40, respectively). We believe these comparisons provide useful context for the remaining gap to BMS.

In the paper, we attribute WASABI’s worse performance on the energy metric to undercoverage of the left tail in LJ-13’s energy distribution. We also notice that our energy distribution is slightly more peaked at the mode, which is consistent with the mode-seeking behavior that reverse-KL-based training can exhibit in practice. The steep Lennard–Jones potential also makes energy particularly sensitive to small differences in particle separations.

Interestingly, BMS exhibits a different coverage pattern on ALA2 and appears to assign more probability mass to low-density regions, whereas WASABI tends to underrepresent certain rare configurations. We believe this difference may be related to the objectives underlying the two approaches. WASABI is motivated by reverse-KL descent, while BMS is connected to a forward-KL objective over path distributions through its Markovian projection formulation. Indeed, the authors of BMS explicitly identify this connection as a possible explanation for their method’s mode-covering behavior (Blessing et al., 2026, Section 4.3). Our observations are consistent with this interpretation and highlight an interesting distinction between the two approaches: broader mode coverage does not necessarily imply more accurate probability allocation, as illustrated by WASABI’s lower reported dihedral JS divergence on ALA2 and its lower W2-TICA metrics.

We hope this clarifies the remaining LJ-13 discrepancy and puts it in the context of WASABI’s overall performance. We believe that extending our denoiser-update formulation to other objectives is a particularly exciting avenue for future work.

W3/Q3: Large system scalability and transferability

We agree with the reviewer that our current experiments establish performance on small molecular benchmarks, and extending WASABI to substantially larger systems remains an important direction for future work. The mathematical formulation applies directly to larger systems, but maintaining sampling accuracy at manageable computational cost requires further investigation. Our ALA2 results, obtained using 4-layer, 64-channel PaiNN backbones compared with the 6-layer, 256-channel backbone used by damped BMS, suggest that architectures tailored to the target energy landscape can help achieve strong performance with compact networks. Exploring such architectural choices, alongside increased model capacity where needed, provides a promising path toward larger systems. Another direction is to evaluate the accuracy and variance of covariance estimation as system size increases and develop more robust estimators to support scaling.

Regarding transferability, the standard GEOM-DRUGS conformer-generation benchmark used by Torsional Diffusion involves learning from conformer datasets and generalizing to unseen molecules. This differs from our focus on learning a sampler for a specified physical energy without requiring equilibrium target samples. We nevertheless view energy-based training across molecular systems, potentially combined with existing conformer data, as a valuable extension.

W3/Q3: End-to-end compute cost and prior initialization

Our training FLOP estimate already includes correction training, denoiser fitting, and generation of fresh replay-buffer samples. Prior preparation is additionally included in the reported energy/force-call budget. These estimates cover the principal training components.

We also re-ran our method with high-temperature MD prior instead of REMD prior to match our FLOP estimate (see response to Q1/W1).

AuthorsOpenReview ↗

Response to reviewer yR4Q

Q1 Fidelity and estimator accuracy

We thank the reviewer for requests of clarification on suggestions of diagnostics.

We have clarified the implemented bridge-time update and covariance-weight schedules in Appendix B.3.

As reviewer requested, we have added a tractable two-dimensional Gaussian study in Appendix B.5, Figure 6, comparing correction accuracy using exact and learned denoisers with full and downweighted covariance contributions. Each fitted correction is evaluated against its corresponding analytic response. In this example, we found that errors decrease with increasing Monte Carlo sample budget, and learned denoisers yield errors close to the exact-denoiser baseline. We also observe larger correction errors at later bridge times; This is consistent with the potentially larger sampling fluctuations associated with the estimator’s t/hₜ and t²/hₜ² factors as the bridge-noise scale decreases. The appendix specifies the experimental settings and distinguishes estimation error from the deliberate change in response introduced by covariance downweighting.

Our tilted-noise construction is an alternative estimator of the same population quantity as the direct Gaussian estimator; under the stated assumptions, both are unbiased. We introduced tilting because the direct estimator exhibited very high variance and prevented stable training in our experiments.

Reviewer_H1H8OpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Wasserstein Sampling via Bridge Interpolants (WASABI), a method for learning diffusion-based samplers for Boltzmann distributions. The central idea is to track the Wasserstein gradient flow of the KL divergence by updating the denoiser of a stochastic interpolant, rather than directly simulating Langevin dynamics. The authors derive a first-order relationship between Wasserstein descent and denoiser updates, and develop regression-based estimators that avoid evaluating the score of the current sampling distribution. The resulting algorithm iteratively updates the diffusion sampler using generated samples and energy gradients.

Experiments on a four-particle double-well system, a Lennard–Jones cluster, and alanine dipeptide demonstrate competitive performance, with particularly promising results on alanine dipeptide in terms of both sampling accuracy and computational cost.

Critical strengths

The main strength of this paper is its interesting and conceptually novel theoretical framework. By deriving how the denoiser (and consequently the drift) of a stochastic interpolant should evolve to track the Wasserstein gradient flow, the authors establish an insightful connection between distributional gradient flows and generative modeling. I find this perspective particularly interesting, as it offers a new way to construct sampling algorithms from prescribed distributional dynamics.

Critical weaknesses
  1. IA central theoretical concern is the introduction of the hyperparameter \lambda_t. According to Theorem 3.1, the denoiser update tracks the Wasserstein gradient flow to first order only when \lambda_t=1 (with a common sufficiently small step size). However, the experiments suggest that smaller values of \lambda_t lead to better performance. When \lambda_t\neq 1, the update no longer corresponds to the derived Wasserstein gradient direction. Although the authors show that the target distribution remains a fixed point, this does not establish convergence toward that fixed point. At least an intuitive explanation of why the modified updates should still progressively approach the target distribution would be important to support the proposed algorithm.

  2. The experimental comparisons focus primarily on recent diffusion- and bridge-based sampling methods, without a broader evaluation against other energy-based generative sampling approaches. Moreover, while WASABI substantially outperforms the reported baselines on alanine dipeptide, it performs noticeably worse than BMS on LJ-13. This raises the question of whether the method has particular advantages for certain molecular systems, rather than providing a general improvement in sampling performance. Finally, alanine dipeptide is still a relatively small molecular system, leaving the scalability of the proposed approach to larger and more complex molecules unclear.

Questions for authors
  1. Is there any theoretical or intuitive justification that the modified denoiser update with \lambda_t\neq 1 still moves the sampling distribution toward the target? For example, can the authors establish conditions under which the KL divergence decreases locally, or explain why convergence should be expected despite the deviation from the Wasserstein gradient direction?

  2. WASABI performs particularly well on alanine dipeptide but is outperformed by BMS on LJ-13. Do the authors have an explanation for this difference? Are there particular properties of the target distribution or molecular system that favor WASABI?

  3. What are the main challenges in extending WASABI to substantially larger molecular systems? In particular, how are the computational cost and variance of the correction estimators expected to scale with dimensionality?

8 comments
AuthorsOpenReview ↗

(1/n) Response to reviewer H1H8

We thank the reviewer for their thoughtful feedback and positive assessment of our theoretical framework. Below, we address covariance downweighting, empirical performance, scalability, and broader comparisons.

W1/Q1: Intuition for covariance-downweighted denoiser updates

Thank you for your feedback! Initially the motivation for downweighting the covariance component was that we measured its variance to be empirically much higher than that of the mean component.

Covariance Weight: Per Theorem 3.1 the wasabi update decomposes into two components Mt\mathcal M_t and Ct\mathcal C_t. Early in our experiments, we noticed that Var(Ct)Var(\mathcal C_t) was much greater than Mt\mathcal M_{t}. We initially decided to model this as having access to a noisy version of Ct\mathcal C_t. Specifically suppose we have access to the Mt\mathcal M_t component of the Wasserstein gradient and to a noisy version of Ct\mathcal C_t

Ct^=Ct+ξ(t)\hat{\mathcal C_t }= \mathcal C_{t} + \xi(t)

Where the noise ξ(t)\xi(t) satisfies E(ξ(t))=0\mathbb{E}(\xi(t)) = 0 and Var(ξ(t))=σt2\text{Var}(\xi(t)) = \sigma_t^2. For each fixed tt, we wanted to find the λt\lambda_t to best approximate the descent direction

L(λt)=E[∣∣Mt+λtCt^−(Mt+Ct)∣∣2]\mathcal L (\lambda_t) = \mathbb{E}[ || \mathcal M_t + \lambda_t \hat{C_t} - (\mathcal M_t + \mathcal C_t) ||^2 ]

Minimizing the loss function above gives the optimal value of λt\lambda_t

λt∗=E[∣∣Ct∣∣2E[∣∣Ct∣∣2]+σt2=SNRt1+SNRt\lambda_t^* = \frac{\mathbb{E}[|| \mathcal C_t ||^2}{\mathbb{E}[||\mathcal C_{t}||^2] + \sigma_t^2} = \frac{\text{SNR}_t}{1 + \text{SNR}_t}

In practice the SNRtSNR_t value of Ct\mathcal C_t is not known and depends on things like the approximation error of the posterior mean as well as the estimation error of Ct\mathcal C_t. However, this does show that shrinking Ct\mathcal C_t can be a good approximation for the true descent direction when the estimate Ct^\hat{\mathcal C_{t}} for Ct\mathcal C_t has high variance. We will include this derivation in the paper, and would like to thank you for the insightful feedback!

AuthorsOpenReview ↗

(2/n) Response to reviewer H1H8

W2/Q2: Performance differences between ALA2 and LJ-13

Thank you for feedback and for the careful review of the experiments. We agree that while BMS achieves a lower LJ-13 energy Wasserstein distance, WASABI remains competitive with AS and ASBS, both strong sampling baselines. Its configuration accuracy is close to ASBS, and its reported mean energy Wasserstein distance is lower than those of both AS and ASBS. Our LJ-13 samples have a slightly higher mean energy than the reference, −42.11 versus −43.13, and a narrower energy distribution, with an average standard deviation of approximately 6.17 versus 7.02. They capture the main part of the distribution but give too little weight to low-energy configurations. This is consistent with the mode-seeking behavior that reverse-KL-based training can exhibit in practice. The steep Lennard–Jones potential also makes energy sensitive to small errors in particle separations, so close agreement in configuration space can coexist with a higher energy Wasserstein distance. We observe similar behavior in ALA2, where WASABI has difficulty representing the rare positive-(\phi) region. At the same time, it achieves a lower overall reported dihedral JS divergence than damped BMS. This illustrates that broader mode coverage does not necessarily mean more accurate probability allocation. The distinction also appears in the theoretical formulations of the methods. WASABI is motivated by reverse-KL descent, while BMS characterizes its ideal Markovian projection through a forward-KL objective over path distributions (Eq. 14, BMS). BMS’s authors themselves conjecture that this connection contributes to their observed mode-covering behavior (§4.3, BMS). This suggests that the choice of objective may contribute to the different coverage patterns. We have included this discussion in the revised manuscript and believe that extending our denoiser-update formulation to other objectives is a particularly exciting avenue for future work. We will include a discussion about this in the experimental section of our paper, thank you for the feedback!

W2/Q3: Scalability to larger molecular systems

We agree with the reviewer that our current experiments establish performance on small molecular benchmarks, and extending WASABI to substantially larger systems remains an important direction for future work. The mathematical formulation applies directly to larger systems, but maintaining sampling accuracy at manageable computational cost requires further investigation. Our ALA2 results, obtained using 4-layer, 64-channel PaiNN backbones compared with the 6-layer, 256-channel backbone used by damped BMS, suggest that architectures tailored to the target energy landscape can help achieve strong performance with compact networks. Exploring such architectural choices, alongside increased model capacity where needed, provides a promising path toward larger systems. Another direction is to evaluate the accuracy and variance of covariance estimation as system size increases and develop more robust estimators to support scaling.

W2: Comparison with broader energy-based sampling approaches

We thank the reviewer for suggesting comparisons with broader sampling approaches. Our primary contribution is a new framework for learning diffusion-based samplers, so we prioritize recent diffusion- and bridge-based methods as the most direct comparisons. As discussed in our related-work section, complementary normalizing-flow approaches employ varied training settings: some learn from simulation datasets and subsequently correct sampling bias, as in Sequential Boltzmann Generators [1], while energy-trained methods such as FAB and TA-BG use internal molecular coordinates instead of atomic coordinates used in our work [2,3]. A systematic comparison across these sampler families would require aligning target distributions, molecule parameterization, evaluation procedures, and computational budgets. We agree that such a study would be valuable and view it as an important extension beyond the present paper’s focus. If you feel like we have missed any important references on previous important work, we’d also be happy to include them!

[1] Tan, C. B., et al. Scalable Equilibrium Sampling with Sequential Boltzmann Generators. ICML, 2025.

[2] Midgley, L. I., et al. Flow Annealed Importance Sampling Bootstrap. ICLR, 2023.

[3] Schopmans, H., and Friederich, P. Temperature-Annealed Boltzmann Generators. ICML, 2025.

I am not sure whether the authors omitted a \lambda_t in the second equation. In addition, it still seems unanswered why \lambda_t<1 can guarantee convergence.

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

WASABI converts energy-based updates into corrections to a diffusion denoiser, without requiring training samples from the target distribution. WASABI achieves better alanine-dipeptide sampling accuracy with lower computational cost.

Critical strengths
  1. The update accounts for both sample movement and changes in which samples contribute to a denoised prediction. Its derivation avoids needing the model's unknown probability density.

  2. The paper gives an exact procedure for drawing the modified noise used in training. It also accounts for how rotating molecular structures into alignment changes the required calculation.

Critical weaknesses
  1. The theory describes an ideal, very small update. Training changes that update and reuses older samples, so the theory does not establish that the implemented procedure consistently moves closer to the target distribution.

  2. The alanine experiment adds separate networks for different types of molecular motion. Without testing this design separately, the accuracy gain cannot be clearly attributed to the proposed update rule.

  3. The reported accuracy uses starting samples from replica-exchange simulation, while the cost estimate charges a cheaper sampling procedure. The paper does not show that the cheaper procedure achieves the same final accuracy.

Questions for authors
  1. With the same network and number of force evaluations, do the sample-weight correction and modified noise improve accuracy? Please provide the update and noise settings used.

  2. Can the reported alanine accuracy be reproduced using the cheaper starting samples included in the cost estimate, including how often rare conformations appear?

  3. How much of the alanine improvement remains without the separate networks for different molecular motions, or when a comparison method receives the same design?

4 comments
AuthorsOpenReview ↗

(1/n) Response to reviewer tsZv

We thank the reviewer for their careful reading, constructive feedback, and recognition of our denoiser-update derivation and tractable estimators. These comments have helped us strengthen both the theoretical presentation and empirical evaluation of our work. Below, we clarify the role of our theoretical results and address the empirical and computational concerns raised.

W1: Role of the theory and convergence of the practical algorithm

This is a really good observation and we agree with the reviewer that the paper should clarify some of these aspects more clearly. We agree that the theory is indeed concerned with an idealized setting and want to clarify that the role of Theorem 3.1 is to guide practical diffusion-sampler design through Wasserstein gradient flow in distribution space. The ideal Wasserstein update requires the unknown score of the current sampling distribution and is therefore not directly tractable. By characterizing the induced first-order denoiser change through its mean and covariance contributions, the theorem guides our construction of a tractable regression update that avoids evaluating the current distribution’s density or score.

We agree that this result does not establish convergence or monotonic KL decrease for the implemented procedure. Finite step size, approximate neural-network fitting, and replay-buffer reuse introduce deviations from the ideal population iteration. However, we believe that the choices of both the covariance weight as well as the replay buffer can be rigorously justified. We provide an explanation for both choices below.

Covariance Weight: Per Theorem 3.1 the wasabi update decomposes into two components Mt\mathcal M_t and Ct\mathcal C_t. Early in our experiments, we noticed that Var(Ct)Var (\mathcal C_t) was much greater than Var(Mt)Var(\mathcal M_{t}). We initially decided to model this as having access to a noisy version of Ct\mathcal C_t. Specifically suppose we have access to the Mt\mathcal M_t component of the Wasserstein gradient and to a noisy version of Ct\mathcal C_{t}

Ct^=Ct+ξ(t)\hat{\mathcal C_t} = \mathcal C_{t} + \xi(t)

Where the noise ξ(t)\xi(t) satisfies E(ξ(t))=0\mathbb{E}(\xi(t)) = 0 and Var(ξ(t))=σt2\text{Var}(\xi(t)) = \sigma_t^2. For each fixed tt, we wanted to find the λt\lambda_t to best approximate the descent direction

L(λt)=E[∣∣Mt+λtC^t−(Mt+Ct)∣∣2]\mathcal{L}(\lambda_t) = \mathbb{E}[ || \mathcal{M}_t + \lambda_t \hat{C}_t - (\mathcal{M}_t + \mathcal{C}_t) ||^2 ]

Minimizing the loss function above gives the optimal value of λt\lambda_t

λt∗=E[∣∣Ct∣∣2E[∣∣Ct∣∣2]+σt2=SNRt1+SNRt\lambda_t^* = \frac{\mathbb{E}[|| \mathcal C_t ||^2}{\mathbb{E}[||\mathcal C_{t}||^2] + \sigma_t^2} = \frac{\text{SNR}_t}{1 + \text{SNR}_t}

In practice the SNRtSNR_t value of Ct\mathcal C_t is not known and depends on things like the approximation error of the posterior mean as well as the estimation error of Ct\mathcal C_t. However, this does show that shrinking Ct\mathcal{C}_t can be a good approximation for the true descent direction when the estimate Ct^\hat{\mathcal C_t} for Ct\mathcal C_t has high variance. We will include this derivation in the paper, and would like to thank you for the insightful feedback!

Using a replay buffer. The replay buffer is used mainly for two reasons. The first is computational efficiency. Without a replay buffer, each outer iteration of WASABI would require generating a large number of fresh samples from the current diffusion model to optimize the objective. Since diffusion sampling is computationally expensive, this would introduce substantial overhead. A replay buffer amortizes this cost by reusing previously generated samples across iterations, trading sample freshness for computational efficiency.

The second motivation is to retain previously visited configurations when the sampler changes rapidly. An aggressive update can reduce the probability of revisiting some regions, even if they were represented in earlier iterations. Keeping older samples in the buffer allows these configurations to remain available during training as the buffer is gradually refreshed. This does not guarantee that coverage is preserved, but it is another practical motivation for using replay.

However, you are correct that this is not exact: samples retained from earlier iterations do not follow the current sampler’s distribution. Replay therefore introduces an additional approximation to the ideal update, which we will clarify in the paper. We also believe that reducing the need for replay, for example, by distilling the sampler into a flow map that generates fresh samples in fewer steps, is an exciting avenue for future work.

Thank you for this feedback. We have revised the Conclusion to clarify the theorem’s role in guiding algorithm design and identify convergence guarantees for the implemented procedure as future work.

AuthorsOpenReview ↗

(n/n) Response to reviwer tsZv

W2/Q3: Contribution of the molecular architecture

We thank the reviewer for raising the contribution of molecular architecture. For ALA2, we view the bond–angle split as a physically motivated contribution complementary to the update rule. It separates motions that change bond lengths and angles from softer conformational motions, allowing the correction networks to represent these different molecular responses separately. Our preliminary tests without this split yielded worse performance. With this design, WASABI achieves strong accuracy using three compact 4-layer, 64-channel PaiNN backbones for both heads, compared with the 6-layer, 256-channel backbone used by damped BMS. We have updated Section 4.3 to make the ALA2-specific architecture explicit. We nevertheless agree that controlled architectural ablations would help distinguish these contributions. A fuller comparison would include adapting BMS and other methods to the same split architecture and evaluating its performance. This requires implementation and validation beyond what is feasible during the rebuttal, and we plan to include it in an expanded version of the paper.

The DW-4 and LJ-13 experiments provide evidence for WASABI’s usefulness without the motion-dependent architecture. On DW-4, WASABI, BMS, and ASBS use matched EGNN architectures and training settings, and WASABI achieves the lowest mean energy Wasserstein distance, providing a direct comparison under matched architectural settings. On LJ-13, WASABI uses four-layer, 64-channel PaiNN backbones without the stiff/soft split, which are shallower and narrower than the five-layer, 128-feature EGNNs used by BMS and ASBS. Its configuration Wasserstein distance is comparable to ASBS, and its mean energy Wasserstein distance is lower than AS and ASBS.

W3/Q2: Cheaper prior generation procedure

We thank the reviewer for pointing this out. In response, we reran the experiments using a sample bank generated directly from a high-temperature molecular dynamics (MD) prior with the number of energy evaluations that were stated in the paper. We have updated the manuscript to reflect this change and included additional evaluation metrics, such as W2-TICA, which assesses agreement between the generated and reference distributions along slow collective coordinates. These metrics complement dihedral-angle-based evaluations by providing a broader assessment of the sampled conformational distributions. We are happy to report that the results are mostly unchanged. We state the median JS as well as the values for the other seeds below.

PriorDihedral JS ↓TICA W2 ↓
REMD0.0267 [0.0242, 0.0273]0.1873 [0.1842, 0.1879]
High-temperature MD0.0235 [0.0187, 0.0251]0.1650 [0.1561, 0.1694]

We hope this clears up any remaining mismatch on the cost accounting between the prior generation procedures.

“With the same network and number of force evaluations, do the sample-weight correction and modified noise improve accuracy? Please provide the update and noise settings used.”

Appendix B.4, Table 5 shows that, under the same architecture and force-evaluation budget at sη=1.00s_{\eta}=1.00, a moderate covariance weight improves mean JS divergence from 0.05950.0595 without the correction (λcov=0\lambda_{\mathrm{cov}}=0) to 0.03650.0365 at λcov=0.40\lambda_{\mathrm{cov}}=0.40.

The tilted-noise estimator was introduced because the direct Gaussian estimator exhibited very high variance in our preliminary experiments. Under the stated assumptions, the two estimators have the same population expectation while the tilted one has better empirical behavior.

The update rule and tilted-noise distribution are given in Eqs. (11) and (15), respectively, with sampling details in Appendices A.4–A.5 and experimental settings in Appendices B.3–B.4.

Reviewer_tsZvOpenReview ↗

Response after author rebuttal

The cheaper prior experiments address the accuracy-cost mismatch, and the additional tests improve the method, but the contribution of the molecular architecture remains unisolated.

Our bond–angle split parameterizes the correction, which is driven by the energy gradient. The BMS network instead regresses a drift onto a bridge-matching target, which is not a force-like quantity, so it is not obvious the same decomposition helps. Nevertheless, we are happy to explore it for BMS.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Response to the reviewers.

We thank the reviewers for their thoughful feedback and constructive comments! We have incorporated the suggested changes into the paper and have highlighted the differences relative to the previous version of the paper with red text. We welcome any questions and suggestions.

Best,

The authors.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.