betaFeedback

WorksMathematics

WASABI: Wasserstein Sampling via Bridge Interpolants

Abstract

Sampling from Boltzmann distributions is a fundamental problem in physics, chemistry, and biology. We introduce WAsserstein SAmpling via Bridge Interpolants (WASABI), a method that trains a diffusion model to generate samples from a target Boltzmann distribution given access only to the energy function. WASABI minimizes the KL divergence between the distribution generated by its diffusion model and the Boltzmann target using Wasserstein gradient flows. This approach builds on a principle underlying classical sampling methods such as Langevin dynamics. However, applying Wasserstein gradient flows to diffusion models is challenging because the updates require the gradient of the model's own log density, which is generally intractable. We derive a practical training algorithm that avoids computing this gradient. Once trained, WASABI can generate new samples without further evaluations of the energy or its gradient. On alanine dipeptide, WASABI outperforms the best reported baselines using an estimated 3.3 times fewer energy evaluations. It also achieves competitive performance on interacting particle system benchmarks.

Introduction

We consider the problem of sampling from a Boltzmann distribution,

π(x)=e−U(x)Z,Z=∫e−U(x) dx,\pi(x) = \frac{e^{-U(x)}}{Z}, \qquad Z = \int e^{-U(x)}\,dx,

where U(x)U(x) is an energy function and ZZ is the normalizing constant. We assume no samples from π\pi are available and have only black-box access to U(x)U(x) and its gradient ∇U(x)\nabla U(x). Many systems in physics, chemistry, and biology are naturally described by Boltzmann distributions. For these systems, samples from π\pi are useful because they allow us to estimate expectations Eπ[f(X)]\mathbb{E}_{\pi}[f(X)]. For appropriate choices of ff, these expectations answer important scientific questions such as how likely a protein is to fold [26, 21], how likely a drug is to bind to a particular protein [5], or, for a material, what its physical properties are and which of its phases is most stable [7, 14].

Classical approaches to Boltzmann sampling include Markov chain Monte Carlo (MCMC; [24]) and molecular dynamics (MD; [7]). These methods start from an initial state x0x_0 and repeatedly update it with small, local changes computed from UU or ∇U\nabla U, producing a sequence of states x0→x1→x2→⋯x_0 \rightarrow x_1 \rightarrow x_2 \rightarrow\cdots. Langevin dynamics is a canonical example of such a classical approach [19]. It updates xtx_t with a small step along −∇U(xt)-\nabla U(x_t), toward more likely states, plus Gaussian noise. Under suitable conditions, the distribution of xtx_t approaches π\pi as t→∞t \rightarrow\infty.

In practice, however, MCMC methods like Langevin dynamics take a long time to converge to π\pi. The reason is that moving from one likely region to another may require passing through unlikely states, and such transitions are rare [30]. Moreover, evaluating UU or ∇U\nabla U at every step can be expensive. In computational chemistry, for example, evaluating the energy of a particular state requires an expensive quantum-mechanical calculation, such as one based on density functional theory [18, 27].

These costs motivate amortized samplers: generative models trained to approximate the Boltzmann distribution. Once trained, these models can produce independent samples cheaply, without evaluating the energy or its gradient. Although training requires such evaluations, their cost is spread, or amortized, across the samples the model later produces. By contrast, obtaining representative samples from π\pi with MCMC or MD can be computationally expensive, as both may require long simulations involving repeated energy or gradient evaluations.

In this work, we introduce WAsserstein SAmpling via Bridge Interpolants (WASABI), an amortized sampler for Boltzmann distributions. WASABI is trained by minimizing the Kullback–Leibler (KL) divergence KL(q∥π)\mathrm{KL}(q\Vert\pi) between its sampling distribution qq and the target π\pi. The idea is to perform gradient descent on this divergence directly in the space of distributions, a procedure known as the Wasserstein gradient flow (WGF; [15]). To implement this flow, WASABI represents qq with a diffusion model, a powerful class of generative models that have been successfully applied to images, video, and protein structures [11, 29, 12, 32]. WASABI repeatedly updates the diffusion model so that its endpoint distribution qq takes small steps along the Wasserstein gradient flow toward π\pi.

Figure 1 illustrates the idea. WASABI trains a diffusion model to follow Wasserstein gradient flow toward π\pi. At each step nn, WASABI seeks to translate a Wasserstein descent step of its current distribution qnq_n into an update of the diffusion model, yielding a new distribution qn+1q_{n+1}. Directly implementing this update would require the gradient of the model’s log density (the model score), which is generally unavailable for diffusion samplers. We show how to perform this update using only samples drawn from qnq_n and target energy gradients ∇U\nabla U, without evaluating the model score. Repeating this procedure yields a diffusion model trained to approximate π\pi, from which samples can be generated without further energy or gradient evaluations.

Overview of WASABI

Figure 1. Overview of WASABI. The endpoint distributions evolve from an initialization q0q_0, distinct from the shared diffusion prior p0p_0. Top: we optimize a sequence of sampling distributions qnq_n by evolve them toward the target π\pi through Wasserstein gradient descent of the KL divergence. Bottom: the sampling distribution qnq_n is parameterized by a diffusion model with a prior p0p_0 for all nn; WASABI translates Wasserstein descent into updates of the diffusion model by updating its denoising representation mˉt(n)\bar{m}_t^{(n)}, gradually moving it toward the target across iterations nn. The final panel shows the ideal diffusion sampler, with denoising representation mˉt∗\bar{m}_t^* associated with π\pi.

We summarize our contributions as follows:

  • Amortizing Langevin dynamics. We introduce a diffusion sampler that amortizes the evolution of distributions under Langevin dynamics. Formally, this evolution is a Wasserstein

gradient flow of the KL divergence to π\pi [15]. Discretizing this flow gives an update that WASABI applies repeatedly to its distribution; in its exact form, the update has π\pi as its fixed point.

  • A tractable training algorithm. We derive a tractable algorithm for applying Wasserstein descent-induced updates to the diffusion model that avoids evaluating the diffusion model score. Each step updates the diffusion model through regression, for which we construct unbiased, empirically low-variance estimators of the training targets.

  • Accurate molecular sampling with fewer energy evaluations. On alanine dipeptide, WASABI reproduces the dominant conformational basins and achieves lower Jensen–Shannon divergence than the best reported baselines [4], using an estimated 30% of their energy/force evaluations.

Background

Diffusion models and bridge interpolants. Diffusion models [11, 29] and related developments [22, 3, 10] provide tools for generative modeling and sampling through continuous-time dynamics. To sample a target distribution p1p_1, we simulate the following SDE driven by a function family mt:Rd→Rdm_t : \mathbb{R}^d \to\mathbb{R}^d:

dXt=mt(Xt)−Xt1−t dt+B dWt,X0∼p0,0≤t<1,(1)\mathrm{d}X_t = \frac{m_t(X_t)-X_t}{1-t}\,\mathrm{d}t + B\,\mathrm{d}W_t, \qquad X_0 \sim p_0,\qquad0 \le t < 1, \tag*{(1)}

where p0p_0 is the source distribution which is easy to sample from, B>0B > 0 is the diffusion coefficient, and WtW_t is standard Brownian motion. The choice of mtm_t that generates the same endpoint distribution p1p_1 need not be unique. We consider a canonical representation of p1p_1 induced by a bridge interpolant:

Yt=(1−t)X~0+tX~1+htZ,0≤t≤1,(2)Y_t = (1-t)\widetilde{X}_0 + t\widetilde{X}_1 + h_t Z,\qquad0 \le t \le1, \tag*{(2)}

where X~0∼p0\widetilde{X}_0 \sim p_0, X~1∼p1\widetilde{X}_1 \sim p_1 are source and endpoint samples and Z∼N(0,Id)Z \sim\mathcal{N}(0,I_d) is the noise, all drawn independently; ht=Bt(1−t)h_t = B\sqrt{t(1-t)} is the bridge noise standard deviation. The associated bridge conditional mean defines the denoising representation below.

Definition 2.1 (Denoising representation of p1p_1). Under the bridge setup in Eq. (2), we call the family

m‾t(y)=E[X~1∣Yt=y],0≤t<1,\overline{m}_t(y) = \mathbb{E}[\widetilde{X}_1 \mid Y_t = y],\qquad0 \le t < 1,

the denoising representation of p1p_1 relative to p0p_0 and BB.

Under suitable regularity conditions, choosing mt=m‾tm_t = \overline{m}_t in Eq. (1) reproduces the bridge marginals (Xt=dYtX_t \overset{d}{=} Y_t) and yields endpoint distribution p1p_1 as t↑1t \uparrow1.

When endpoint samples X~1∼p1\widetilde{X}_1 \sim p_1 are available, a neural denoiser mω(y,t)m_\omega(y,t) is trained using the denoising regression objective, whose population minimizer is m‾t(y)\overline{m}_t(y):

min⁡ωEt,X~0,X~1,Z[∥mω(Yt,t)−X~1∥2].(3)\min_\omega\mathbb{E}_{t,\widetilde{X}_0,\widetilde{X}_1,Z}\left[\left\|m_\omega(Y_t,t)-\widetilde{X}_1\right\|^2\right]. \tag*{(3)}

Wasserstein gradient flows. Let P2(Rd)\mathcal{P}_2(\mathbb{R}^d) denote the probability measures on Rd\mathbb{R}^d with finite second moment. In Euclidean space, the gradient describes how a function changes when its input is displaced. Here, we consider a functional F:P2(Rd)→R\mathcal{F}:\mathcal{P}_2(\mathbb{R}^d)\to\mathbb{R} that assigns a scalar objective value to a probability distribution, e.g. a divergence to some target distribution. Its gradient likewise describes how the objective value F\mathcal{F} changes when that distribution is displaced. We displace a distribution qq in the direction of a vector field vv by moving each point xx to x+ϵv(x)x+\epsilon v(x), where ϵ>0\epsilon>0 is a small step size, which gives qϵν=(id⁡+ϵv)#qq^\nu_\epsilon=(\operatorname{id}+\epsilon v)_\#q, where id⁡\operatorname{id} is the identity map and #\# denotes the pushforward [28].3 This notion of displacement allows us to define the Wasserstein gradient of F\mathcal{F} at qq, denoted ∇W2F[q]\nabla_{W_2}\mathcal{F}[q], as the vector field satisfying

ddϵF(qϵν)∣ϵ=0=EX∼q[∇W2F[q](X)⋅v(X)](4)\left.\frac{\mathrm{d}}{\mathrm{d}\epsilon}\mathcal{F}(q^\nu_\epsilon)\right|_{\epsilon=0} = \mathbb{E}_{X\sim q}\left[\nabla_{W_2}\mathcal{F}[q](X)\cdot v(X)\right] \tag*{(4)}

for every sufficiently regular velocity field vv [15]. This identity expresses the directional derivative of F\mathcal{F} as an L2(q)L^2(q) inner product; consequently, the negative Wasserstein gradient gives the direction of steepest descent among velocity fields vv with the same L2(q)L^2(q)-norm Eq∥v∥2\mathbb{E}_q\|v\|^2.

We say that a sufficiently regular curve (qτ)τ≥0(q_\tau)_{\tau\ge0} follows 2-Wasserstein gradient flow if, to first order,

qτ+ϵ=(id⁡−ϵ∇W2F[qτ])#qτ+o(ϵ),(5)q_{\tau+\epsilon} = \left(\operatorname{id} - \epsilon\nabla W_2 \mathcal{F}[q_\tau]\right)_{\#}q_\tau+ o(\epsilon), \tag*{(5)}

where the expansion is understood weakly, that is, after integration against smooth test functions. Equivalently, when qτq_\tau admits a density, its evolution satisfies the continuity equation

∂τqτ=∇⋅(qτ∇W2F[qτ]).(6)\partial_\tau q_\tau= \nabla\cdot\left(q_\tau\nabla W_2 \mathcal{F}[q_\tau]\right). \tag*{(6)}

For sampling the Boltzmann target π(x)∝e−U(x)\pi(x) \propto e^{-U(x)}, consider the reverse KL objective F[q]=KL⁡(q∥π)\mathcal{F}[q] = \operatorname{KL}(q \mathbin\Vert\pi). Its Wasserstein gradient is

References

  1. [1]Tara Akhound-Sadegh, Jarrid Rector-Brooks, Joey Bose, Sarthak Mittal, Pablo Lemos, Cheng-Hao Liu, Marcin Sendera, Siamak Ravanbakhsh, Gauthier Gidel, Yoshua Bengio, Nikolay Malkin, and Alexander Tong. Iterated denoising energy matching for sampling from boltzmann densities. In Proceedings of the 41st International Conference on Machine Learning, volume 235 of Proceedings of Machine Learning Research, pp. 760–786. PMLR, 21–27 Jul 2024. URL https://proceedings.mlr.press/v235/akhound-sadegh24a.html.
  2. [2]Tara Akhound-Sadegh, Jungyoon Lee, Joey Bose, Valentin De Bortoli, Arnaud Doucet, Michael M. Bronstein, Dominique Beaini, Siamak Ravanbakhsh, Kirill Neklyudov, and Alexander Tong. Progressive inference-time annealing of diffusion models for sampling from boltzmann densities. In The Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025. URL https://openreview.net/forum?id=vf2GHcxzMV.
  3. [3]Michael Albergo, Nicholas M Boffi, and Eric Vanden-Eijnden. Stochastic interpolants: A unifying framework for flows and diffusions. Journal of Machine Learning Research, 26(209):1–80, 2025.
  4. [4]Denis Blessing, Lorenz Richter, Julius Berner, Egor Malitsky, and Gerhard Neumann. Bridge matching sampler: Scalable sampling via generalized fixed-point diffusion matching. In Forty-third International Conference on Machine Learning, 2026. URL https://openreview.net/forum?id=gUbNNUuNJo. Code: https://github.com/DenisBless/BridgeMatchingSampler.arxiv.org/abs/2603.00530
  5. [5]I Buch, M J Harvey, T Giorgino, D P Anderson, and G De Fabritiis. High-throughput all-atom molecular dynamics simulations using distributed computing. J Chem Inf Model, 50(3):397–403, Mar 2010. ISSN 1549-960X (Electronic); 1549-9596 (Linking). doi: 10.1021/ci900455r.DOI
  6. [6]Mouyang Cheng, Denis Blessing, Botao Yu, Gerhard Neumann, Mingda Li, Carles Domingo-Enrich, and Yuanqi Du. Atlas: A foundation neural sampler for amorphous materials, 2026. URL https://arxiv.org/abs/2607.19198.
  7. [7]Daan Frenkel and Berend Smit. Understanding Molecular Simulation: From Algorithms to Applications. Academic Press, San Diego, CA, 2 edition, 2002.DOI
  8. [8]Aaron J Havens, Benjamin Kurt Miller, Bing Yan, Carles Domingo-Enrich, Anuroop Sriram, Daniel S. Levine, Brandon M Wood, Bin Hu, Brandon Amos, Brian Karrer, Xiang Fu, Guan-Horn Liu, and Ricky T. Q. Chen. Adjoint sampling: Highly scalable diffusion samplers via adjoint matching. In Proceedings of the 42nd International Conference on Machine Learning, volume 267 of Proceedings of Machine Learning Research, pp. 22204–22237. PMLR, 13–19 Jul 2025. URL https://proceedings.mlr.press/v267/havens25a.html.arxiv.org/abs/2504.11713
  9. [9]Aaron J Havens, Brian Karrer, and Neta Shaul. Flow sampling : Learning to sample from unnormalized densities via denoising conditional processes. In Forty-third International Conference on Machine Learning, 2026. URL https://openreview.net/forum?id=YlcyOMTPN1.arxiv.org/abs/2605.03984
  10. [10]Jeremy Heng, Valentin De Bortoli, Arnaud Doucet, and James Thornton. Simulating diffusion bridges with score matching. Biometrika, 112(4):asaf048, 2025.
  11. [11]Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840–6851, 2020.arxiv.org/abs/2006.11239
  12. [12]Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David Fleet. Video diffusion models. Advances in neural information processing systems, 35:8633–8646, 2022.
  13. [13]John David Jackson. Classical Electrodynamics. Wiley, New York, NY, 3rd edition, 1999. ISBN 9780471309321. URL https://cds.cern.ch/record/490457.
  14. [14]C. Jarzynski. Targeted free energy perturbation. Physical Review E, 65(4):046122, April 2002. doi: 10.1103/PhysRevE.65.046122. URL https://doi.org/10.1103/PhysRevE.65.046122.
  15. [15]Richard Jordan, David Kinderlehrer, and Felix Otto. The variational formulation of the fokker–planck equation. SIAM Journal on Mathematical Analysis, 29(1):1–17, 1998. doi: 10.1137/S0036141096303359. URL https://doi.org/10.1137/S0036141096303359.
  16. [16]Leon Klein and Frank Noé. Transferable Boltzmann generators. In Advances in Neural Information Processing Systems, volume 37, 2024. URL https://proceedings.neurips.cc/paper_files/paper/2024/hash/50354a09f5798e188079e236f437e522-Abstract-Conference.html.arxiv.org/abs/2406.14426
  17. [17]Jonas Köhler, Leon Klein, and Frank Noe. Equivariant flows: Exact likelihood generative learning for symmetric densities. In Proceedings of the 37th International Conference on Machine Learning, volume 119 of Proceedings of Machine Learning Research, pp. 5361–5370. PMLR, 13–18 Jul 2020. URL https://proceedings.mlr.press/v119/kohler20a.html.arxiv.org/abs/2006.02425
  18. [18]W. Kohn and L. J. Sham. Self-consistent equations including exchange and correlation effects. Physical Review, 140(4A):A1133–A1138, 1965. doi: 10.1103/PhysRev.140.A1133.DOI
  19. [19]Benedict Leimkuhler and Charles Matthews. Molecular Dynamics: With Deterministic and Stochastic Numerical Methods. Springer, Cham, 2015.DOI
  20. [20]Zeyang Li, Yunan Wang, Paolo Giardetta, and Navid Azizan. Newton matching for generative modeling: A unified framework for fine-tuning and sampling. arXiv preprint arXiv:2609.05727, 2026.arxiv.org/abs/2609.05727
  21. [21]Kresten Lindorff-Larsen, Stefano Piana, Ron O. Dror, and David E. Shaw. How fast-folding proteins fold. Science, 334(6055):517–520, 2011. doi: 10.1126/science.1208351. URL https://www.science.org/doi/abs/10.1126/science.1208351.
  22. [22]Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matthew Le. Flow matching for generative modeling. In The Eleventh International Conference on Learning Representations, 2023. URL https://openreview.net/forum?id=PqvMRDCJT9t.arxiv.org/abs/2210.02747
  23. [23]Guan-Horng Liu, Jaemoo Choi, Yongxin Chen, Benjamin Kurt Miller, and Ricky T. Q. Chen. Adjoint schrodinger bridge sampler. In The Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025. URL https://openreview.net/forum?id=rMhQB1hh4c.arxiv.org/abs/2506.22565
  24. [24]Jun S. Liu. Monte Carlo Strategies in Scientific Computing. Springer, New York, NY, 2001.
  25. [25]Laurence Illing Midgley, Vincent Stimper, Gregor N. C. Simm, Bernhard Schölkopf, and José Miguel Hernández-Lobato. Flow annealed importance sampling bootstrap. In The Eleventh International Conference on Learning Representations, 2023. URL https://openreview.net/forum?id=XCTVFJwS9LJ.arxiv.org/abs/2208.01893
  26. [26]Frank Noé, Christof Schütte, Eric Vanden-Eijnden, Lothar Reich, and Thomas R. Weikl. Constructing the equilibrium ensemble of folding pathways from short off-equilibrium simulations. Proceedings of the National Academy of Sciences, 106(45):19011–19016, 2009.DOI
  27. [27]Frank Noé, Alexandre Tkatchenko, Klaus-Robert Müller, and Cecilia Clementi. Machine learning for molecular simulation. Annual Review of Physical Chemistry, 71 (Volume 71, 2020):361–390, 2020. ISSN 1545-1593. doi: https://doi.org/10.1146/annurev-physchem-042018-052331. URL https://www.annualreviews.org/content/journals/10.1146/annurev-physchem-042018-052331.
  28. [28]F. Santambrogio. Optimal Transport for Applied Mathematicians: Calculus of Variations, PDEs, and Modeling. Progress in Nonlinear Differential Equations and Their Applications. Springer International Publishing, 2015. ISBN 9783319208282. URL https://books.google.com/books?id=UOHH CgAAQBAJ.DOI
  29. [29]Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2021. URL https://openreview.net/forum?id=PxTIG12RRHS.
  30. [30]Luke Tierney. Markov chains for exploring posterior distributions. Annals of Statistics, 22:1701–1728, 1994. URL https://api.semanticscholar.org/CorpusID:13369978.DOI
  31. [31]Francisco Vargas, Will Sussman Grathwohl, and Arnaud Doucet. Denoising diffusion samplers. In The Eleventh International Conference on Learning Representations, 2023. URL https://openreview.net/forum?id=8pvnfTAbul1f.arxiv.org/abs/2302.13834
  32. [32]Joseph L Watson, David Juergens, Nathaniel R Bennett, Brian L Trippe, Jason Yim, Helen E Eisenach, Woody Ahern, Andrew J Borst, Robert J Ragotte, Lukas F M Filles, et al. De novo design of protein structure and function with rfdiffusion. Nature, 620(7976):1089–1100, 2023.DOI
  33. [33]Qinsheng Zhang and Yongxin Chen. Path integral sampler: A stochastic control approach for sampling. In International Conference on Learning Representations, 2022. URL https://openreview.net/forum?id=_uCb2ynRu7Y.arxiv.org/abs/2111.15141

Paper details

vlnc:W_fdpd5kg6
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 12

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents