acceptodds
Under review as a conference paper at ICLR 2027

Amortised inference through one-step implicit sampling

10 views

Abstract

Efficient sampling from Bayesian posterior distributions in high dimensions remains a hard problem, despite advances in theory and algorithms. Multiple steps are usually required to draw unbiased samples, for example, iterations of an MCMC algorithm or integration steps in a diffusion-based amortised sampler. However, scalable one-step amortised inference remains a challenging open problem. In this work, we build upon recent progress in data-driven generative modelling to offer a possible solution for one-step amortised sampling. Our one-step implicit sampler (OSIS) uses short-time MCMC dynamics to define a local target for the evolution of an implicitly modelled distribution. In the asymptotic limit, the modelled distribution is a fixed point of the MCMC kernel and thus samples from the target. We analyse the convergence properties of OSIS and show connections with other generative modelling and sampling algorithms. We study various design choices and show that OSIS performs well compared to multi-step amortised samplers and non-amortised MCMC baselines while being far more efficient at sampling time.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
one-step samplers, amortised inference, sampling, MCMC, optimal transport, Wasserstein gradient flows, drifting models

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 2 reviews, above 27% of the others.
Reviewer_TfVROpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes a learning-based method for one-step sampling.

The key idea is the target distribution is a fixed point for its MCMC kernel. Therefore, given a kernel K, one can optimize the distributional distance between KρθK\rho_\theta and ρθ\rho_\theta.

Based on the insight, the authors test different choices of the kernel and the loss function, and compare their results with existing methods on both simulated examples and real datasets.

Critical strengths
  1. I found the idea of treating target distribution as a fixed point very elegant. The corresponding learning algorithm is also clean and simple.

  2. Some of the empirical results seem promising.

  3. The paper is very well-written.

Critical weaknesses
  1. I do not agree with some of the claims, such as "OSIS performs similarly – higher mean reward, lower FID – to the diffusion-based amortised baseline (TB+MALA) while requiring only one function evaluation per sample" (Table 4 shows TB + MALA is clearly better). Also, Table 1 shows the OSIS is clearly much worse than many samplers in terms of sampling accuracy.

  2. OSIS has systematic bias (as training is never perfect). Therefore I think the paper is better positioned as a very fast sampler while having bias. The method should also compare with other inexact samplers such as ULA.

  3. I suspect there is a tradeoff between speed vs the complexity of the target distribution. For very complex multimodal distribution, I suspect the current simple training mechanism will not help escaping local modes. Therefore annealing or diffusion-based method might be unavoidable. I did not see much discussion on these potential failure cases.

Questions for authors
  1. Can you discuss more on the initialization strategy? I think it's important for multimodal targets.

  2. Can you comment on what concrete applications that can benefit from your method? I suppose the systematic bias will preclude most Monte Carlo estimation tasks. Also the simplicity (which is overall a good thing!) might also make the method not competitive for very complex distributions. I'd like to know the sweetspot of the method.

I'd be happy to increase the score if the weaknesses and questions are sufficiently discussed by the authors during the rebuttal.

3 comments
AuthorsOpenReview ↗

Response to Reviewer TfVR

We thank the reviewer for the constructive review and for offering to reconsider the score.

(W1) OSIS performance

Please find some comments regarding the claims and efficiency comparison of OSIS with the baselines in Tables G1 and G2 of the general response. Also, please see the updated Table 4 and Table 1 for more recent baselines and updated results. Table 4 shows that the difference between OSIS and TB+MALA lies in the trade-off between average log-reward and FID scores. The higher FID scores of OSIS can be attributed to slightly worse mode coverage; however, OSIS is capable of sampling images with a much higher log-reward. At the same time, TB+MALA is quite underwhelming in terms of log-reward, and OSIS FID can be improved with a short MCMC refinement.

It is worth pointing out that TB+MALA is a much more complex algorithm that requires simulating trajectories and calculating trajectory-level divergences, while OSIS is much simpler. Please see the discussion on cost comparison in the general response.

(W2) Systematic bias

Regarding the bias, any learnt component, be it a diffusion sampler or a learnt one-step amortised sampler, introduces some approximation error. However, because OSIS is implicit, we cannot apply an importance-sampling correction to debias its samples. An alternative approach is to use the generated samples to initialise π\pi-invariant MALA or PT-MALA chains. To reduce this systemic bias, we also used MALA rather than ULA in the main experiment to avoid introducing further bias during training. This discussion will be added to the manuscript revision.

(W3-Q1) Speed / complexity tradeoff

We note that the speed versus complexity trade-off holds true for virtually any sampling algorithm, be it a classic MCMC or an amortised sampler. Training samplers for high-dimensional distributions with complex geometries is difficult and requires additional techniques such as replay buffers or tempering. However, when done right, it is possible to train an efficient and scalable amortised sampler for high-dimensional distributions. This is exactly why we discuss several techniques in Section 5 and Appendix C.1 (MALA step-size adaptation), as well as Appendix A.2 and Algorithm 3, which help us stabilise the training and scale the algorithm. We also provide here an extended discussion on practical considerations below, which we will include in the camera-ready version.

Replay buffer. A replay buffer is a standard technique in reinforcement learning and diffusion-based sampling (e.g., Sendera et al. (2024)) that stores samples from previous iterations to help prevent mode collapse. In this work, we use a replay buffer to store samples generated by the model at the previous outer step. These samples are then resampled with replacement during the inner iterations.

Inner and outer iterations. We found empirically that performing several gradient updates on samples obtained from a short MCMC run at each step stabilises training and improves convergence, while also significantly reducing the number of energy evaluations. This keeps the generator's target distribution fixed during the inner loop and lets the generator more accurately approximate the distribution obtained after each MCMC update.

Initialisation strategy. We do not use any sophisticated initialisation strategy beyond matching the output scale of the generator to the scale of the target distribution. This prevents MCMC from immediately collapsing to a single mode. This is a common practice when sampling with MCMC algorithms.

Choice of neural network. Modelling a multimodal distribution with a single forward pass through a neural network is complicated, as the network may struggle to separate the modes, cf. the famous ‘string’ behaviour seen in normalising flows. Although we use a simple ResNet-based architecture for all our experiments, using a better network may help improve the quality of generated samples.

(Q2) OSIS applicability

We want to highlight two key benefits of OSIS: (1) it is a one-step sampler and (2) it does not require energy evaluations at sampling time, which makes sampling very cheap. We therefore believe OSIS has the potential to be applied to any problem that requires fast sampling. For example, it can serve as an initial distribution for any MCMC method. In practice, MCMC often requires a long burn-in before it samples from the target distribution, and this burn-in must be repeated every time the chain is run from scratch. OSIS can amortise this cost: it is trained once, and its samples are then used to initialise the MCMC chains, which considerably shortens convergence time. Finally, we highlight that OSIS is a much simpler method that surprisingly performs on par with much more sophisticated methods like diffusion samplers or amortised SMC methods. The simplicity and flexibility of our method pave the way for improvement and scaling of similar methods in the future.

thanks for the additional results and careful response!

I understand (and appreciate) it's one-step, very fast, and may be used as high-accuracy algorithm such as Metropolis-Hastings's initialization (though one can argue other neural samplers can also do these). At the same time, it has all the issues just like any other neural samplers, such as intrinsic bias, and mode coverage, and so on.

Therefore, to me, the main evaluation criterion is whether there's a clear and practically meaningful sweetspot, instead of outperforming other methods on all tasks. After reading the rebuttal, I found it still somewhat unclear about the claimed advantage rather than being another neural sampler.

I still find the framework elegant. Nevertheless the concerns seem substantial and therefore I'll maintain my score.

We sincerely appreciate your response and your engagement in the discussion.

We would first like to emphasise that OSIS inherits some limitations of the MCMC kernel it distils. This is discussed in the manuscript, and we will make it more explicit.

The practical sweet spot, which we will also emphasise more clearly, is fast sampling without energy evaluations at sampling time. OSIS is therefore relevant to applications that require fast inference and can tolerate some approximation error.

We further emphasise that OSIS is the first sampler of its kind to be evaluated on large-scale targets. We also develop a theoretical perspective based on Wasserstein gradient flows and establish connections to existing generative models, such as drifting models. We believe these contributions represent a meaningful advance towards new and more efficient approaches to sampling.

We would also appreciate clarification on which of your original concerns were addressed by our response and which remain unresolved, beyond the need to clarify this practical advantage. This would help us address your remaining concerns more precisely.

We would have welcomed further discussion of these points, and we sincerely thank you for your thoughtful feedback.

Sincerely,

The authors

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

This paper proposes OSIS, a method for training a one-step implicit generator to sample from an unnormalised target distribution without access to target samples. The method minimises a discrepancy between the generator's distribution and its refinement through a target-invariant MCMC kernel, with gradients stopped through the refinement. Under suitable assumptions, the target distribution is the unique fixed point of this kernel. The paper analyses convergence of an idealised formulation and evaluates the method on synthetic targets, NICE-based image densities, and posterior sampling under pretrained generative priors.

Critical strengths
  1. Theoretical contribution. I appreciate the convergence analysis and the connection to Wasserstein gradient flows. The results establish convergence of the idealised algorithm under strong log-concavity, with extensions under a log-Sobolev inequality and a curvature condition. This provides a useful theoretical foundation for the proposed approach, subject to the stated assumptions.

  2. Efficient one-step sampling. The paper addresses the important problem of amortising expensive sampling procedures into a single generator evaluation. Training an implicit generator through sample-based discrepancies, without requiring a tractable generator density, is appealing.

Critical weaknesses
  1. The practical loss needs a clearer explanation in the main text. Equation (2) and Algorithm 1 describe the objective at the distribution level, and Section 5 identifies the empirical discrepancies used. However, understanding how the loss is actually computed still requires consulting Appendix C.2 and Algorithm 3. A brief empirical loss formula, together with an explanation of how the generated and MCMC-refined batches are compared and where gradients are stopped, would substantially improve clarity. This is particularly relevant because the theoretical analysis uses squared Wasserstein distance, whereas the main synthetic experiments use sliced Wasserstein distance.

  2. The discussion of existing one-step neural samplers is incomplete. He et al. [1] directly study training one-step implicit generators from unnormalised densities using reverse diffusive KL divergence. This work should be cited and discussed, including differences in the objectives, use of MCMC, and computational requirements. An empirical comparison would also help establish the practical contribution relative to an existing method addressing the same sampling setting.

  3. The evidence for scalability to difficult Boltzmann sampling problems remains limited. The conventional synthetic benchmarks reach only 10 dimensions for 40GMM and 64 dimensions for ManyWell. The NICE experiments provide higher-dimensional density-sampling benchmarks, with dimensions 196 and 784, but they do not establish how performance changes with dimension and multimodal difficulty. FAB also performs competitively on these targets and achieves better results on NICE-MNIST, leaving the claimed scalability advantage insufficiently established.

    Furthermore, the outsourced image experiments in Section 6.2 operate in latent spaces of dimension 16 and 64. The dimensionality of the decoded images should therefore be distinguished from that of the sampling problem. A controlled dimension sweep on the existing GMM benchmark, including 50 dimensions or higher, would provide more direct evidence. A molecular benchmark such as alanine dipeptide would provide complementary evidence of practical relevance.

  4. The computational comparison does not establish the overall amortisation benefit. Tables 1–2 report neural-network and energy evaluations per generated sample, and Figure 3 measures sampling time. These demonstrate the inference-cost advantage. However, training cost is also central when target-energy evaluations are expensive. Appendix D provides training configurations, but the paper still lacks a comparison of total training energy evaluations, energy-gradient evaluations, and wall-clock time across samplers.

    Reporting sampling quality against these training costs would make the comparison more informative. It would also help to estimate how many generated samples are needed to recover the initial training expense. Comparisons with recent methods designed for efficient training, particularly Adjoint Sampling [2], are relevant to these claims. Adjoint Sampling is already cited but is absent from the experiments. Bridge Matching Sampler [3] would provide additional context for the scalability claims. Comparisons should account for both training cost and the different inference costs of one-step and multistep methods.


[1] Jiajun He, Wenlin Chen, Mingtian Zhang, David Barber, and José Miguel Hernández-Lobato. Training Neural Samplers with Reverse Diffusive KL Divergence. AISTATS, 2025. PMLR 258:5167–5175. arXiv:2410.12456.

[2] Aaron Havens et al. Adjoint Sampling: Highly Scalable Diffusion Samplers via Adjoint Matching. ICML, 2025. arXiv:2504.11713.

[3] Denis Blessing, Lorenz Richter, Julius Berner, Egor Malitskiy, and Gerhard Neumann. Bridge Matching Sampler: Scalable Sampling via Generalized Fixed-Point Diffusion Matching. ICML, 2026. arXiv:2603.00530.

Questions for authors

Please find in the weakness section

3 comments
AuthorsOpenReview ↗

Response to Reviewer mNZh

We would like to thank the reviewer for the detailed comments on our paper and the insightful suggestions regarding ways to improve it.

(W1) Practical loss clarification

We are grateful to the reviewer for pointing out the lack of clarity in the discussion of the practical loss. We will amend the manuscript and clarify the loss used in practice. Here we detail how the loss is computed.

At each outer step, we draw z1,…,zB∼i.i.d.N(0,I)z_1,\ldots,z_B\overset{\mathrm{i.i.d.}}{\sim}\mathcal N(0,I) and generate refined samples yi∼Kβ ⁣(⋅∣sg⁡[Gθ(zi)])y_i\sim K_\beta\!\left(\cdot\mid\operatorname{sg}[G_\theta(z_i)]\right), initialising an MCMC chain with samples from GθG_\theta, to which stopgrad (sg⁡[⋅]\operatorname{sg}[\cdot]) is applied. We then take MM (inner steps) gradient updates of θ\theta using the pairs (zi,yi)i=1B(z_i, y_i)_{i=1}^{B}, without running any new MCMC.

Computation of sliced Wasserstein. At the mm-th update, we subsample a batch of B~≤B\tilde B \le B pairs (zj,yj)j=1B~(z_j, y_j)_{j=1}^{\tilde B} (random subset when resampling is enabled), draw fresh directions ω1(m),…,ωP(m)\omega^{(m)}_1,\ldots,\omega^{(m)}_P uniformly on the unit sphere, with PP the number of SWD directions, and take one AdamW step on

L^m(θ)=1P∑p=1P1B~∑r=1B~(⟨ωp(m),Gθ(z)⟩(r)−⟨ωp(m),y⟩(r))2,\widehat{\mathcal L}_m(\theta)= \frac{1}{P}\sum_{p=1}^{P} \frac{1}{\tilde B}\sum_{r=1}^{\tilde B} \Big( \big\langle\omega^{(m)}_p,G_\theta(z)\big\rangle_{(r)} - \big\langle\omega^{(m)}_p,y\big\rangle_{(r)} \Big)^2,

where ⟨ω,Gθ(z)⟩(r)\langle\omega,G_\theta(z)\rangle_{(r)} is the rr-th smallest value of {⟨ω,Gθ(zj)⟩}j=1B~\{\langle\omega,G_\theta(z_j)\rangle\}_{j=1}^{\tilde B}, and likewise for yy.

How the batches are compared. For each direction, sorting gives the exact one-dimensional optimal coupling. This allows scalable computation of SWD.

Where gradients are stopped. The stop-gradient operation is on the copy of the generator output that is used to initialise the chain (sg⁡[⋅]\operatorname{sg}[\cdot] in the manuscript’s Eq.(2)), so the refined samples yy are constants with respect to θ\theta, and the gradient flows to unrefined generator samples.

SWD and the W2\mathcal{W}_2 theory. While the Wasserstein space convergence theory is built for W2\mathcal{W}_2 discrepancy, the fixed point iteration view of OSIS for a π\pi-invariant MCMC kernel KK applies in both cases. In our work, we use SWD primarily for its scalability. It substantially reduces the cost of comparing empirical distributions, enabling faster training and more tuning within a fixed compute budget. In our GMM runs, training with exact OT took about 60 times longer. See also the response to W2 for Reviewer oNmA.

(W2) Related one-step neural samplers

We would like to thank the reviewer for pointing out this recent work on one step sampling. We will include the discussion of this work to the manuscript and we will add the evaluations of the suggested method to Table 1.

However, we empirically found that the algorithm collapses on GMM40 in 10D.

(W3) Scalability

We thank the reviewer for this suggestion and address it in three parts.

First, our benchmark suite follows a substantial body of prior literature (PIS, DIS, SubTB+SMC, etc.), which generally evaluates on the same targets, or simpler ones. This allows direct comparison with established methods.

Second, as the reviewer acknowledges, the NICE targets (196D and 784D) are already high-dimensional benchmarks. Alanine dipeptide (ALDP) has 66 dimensions in Cartesian coordinates (60 in internal coordinates). That is comparable to targets we already consider, such as ManyWell, and far smaller than the NICE targets. It is therefore unclear what additional insight ALDP would provide about scalability. Our existing experiments also already include targets with complex geometry.

Third, whilst a 40-mode GMM in 50D with narrow modes is a useful stress test for mode coverage, it does not correspond to a realistic application, so performance on it says little about a sampler's practical value. In contrast, Bayesian inference under a pretrained generative model prior with a complex reward is both challenging and practically relevant (see the various applications to images, proteins, etc. in Venkatraman et al. (2026)) in many domains where pretrained priors exist and need to be conditioned or constrained.


(W4) Training considerations

Please find the requested measurements of neural net and energy evaluations, as well as the related discussion, in Tables G1 and G2.

For comparisons to other algorithms, please see the the updated Table 1 of the manuscript, where we have included Adjoint Sampler and Bridge Matching Sampler (with DiKL to follow once training finishes).

Thanks the authors for clarifying my concerns and most of them are resolved. The remaining concerns are:

  1. Even ALDP is lower dimensional than the NICE target, I'm not convinced if NICE would be a more difficult problem. At least, the ALDP benchmark would have better impact in the literature of neural sampler because it is relatively more realistic than the NICE; I admit the experiments for ALDP might not be able to be done during the short rebuttal period, but I suggest including it or larger toys (eg ALA-4, ALA-8 etc) in the future revision.

  2. I'm more concerned on the energy evaluation comparison. The latest table shows that OSIS is using as many as energy evaluations as FAB, or even more, while FAB is already left behind in the energy-call-efficiency compared to latest methods.

Overall, I would like to maintain my rating.

We sincerely thank the reviewer for the timely reply.

To further address your remaining concerns, we evaluated OSIS on the ALDP target, following the implementation of Flow Annealed Bootstrap. We agree that ALDP, as a complementary evaluation, is relevant, particularly to the protein design field. We would still argue that it targets a different application from NICE or sampling in the latent space of generative priors, but both have their own value.

We report the metrics and visualisation of results below. The core of OSIS is to distil any MCMC refinement chain. To converge efficiently on ALDP, we chose a well-suited MCMC refinement for the task, which is standard in molecular simulations. It consists of a generalised Hamiltonian Monte Carlo scheme (momentum not fully refreshed but partially renoised) combined with a bound rotation MH step (Vitalis & Pappu, 2009).

MethodKL φ (×10⁻³)KL ψ (×10⁻³)KL γ₁ (×10⁻³)KL γ₂ (×10⁻³)KL γ₃ (×10⁻³)W₂(φ, ψ)KL (φ, ψ) (×10⁻³)
PIS (Zhang & Chen, 2022)†50 ± 30380 ± 4905610 ± 12404490 ± 304600 ± 301.27 ± 1.19-
DDS (Vargas et al., 2023)†30 ± 10160 ± 702440 ± 96030 ± 0.00‡30 ± 0.000.68 ± 0.09-
AS (Havens et al., 2025)†90 ± 9040 ± 40170 ± 170560 ± 90510 ± 600.65 ± 0.52-
ASBS (Liu et al., 2025)†20 ± 0.00‡10 ± 0.0030 ± 1020 ± 0.0020 ± 0.000.25 ± 0.01-
FAB (Midgley et al., 2023)*------26.50 ± 1.20
OSIS (GHMC)93 ± 921 ± 810.5 ± 1.2194 ± 2120.7 ± 2.10.36 ± 0.09156.27 ± 8.74
OSIS (GHMC) + 5 MCMC steps1.62 ± 0.170.51 ± 0.110.27 ± 0.013.58 ± 0.220.42 ± 0.070.13 ± 0.026.22 ± 0.31
Reference floor (MD vs MD)0.19 ± 0.030.23 ± 0.030.20 ± 0.010.27 ± 0.050.19 ± 0.020.11 ± 0.014.74 ± 0.19

†: Metrics extracted from Liu et al., 2025. *: Metrics extracted from Midgley et al., 2023.

These results show that although FAB is a strong sampler on its own, OSIS generally performs comparably to or better than other fully amortised neural samplers from past work, and even better when refined by a few MCMC steps.

Additional visualisations of the samples and energies are available at this link: [link]. We will include these results and further analysis in the final version.

We hope this resolves the reviewer's remaining concerns.

Vitalis & Pappu, “Methods for Monte Carlo simulations of biomacromolecules”, 2009

Comments on the paper

4 comments
AuthorsOpenReview ↗

Response to all reviewers

We would like to thank the reviewers for the detailed and insightful feedback. All reviewers asked questions about the training cost and time for OSIS compared to other baselines, as well as a more detailed description of the conclusions. We provide the answers to these questions in a general response. We later refer to this response in the answers to each reviewer.

Manuscript corrections

The following will be corrected in the manuscript to improve its clarity.

  1. With the paper's settings, the off-policy diffusion sampling algorithms SubTB+SMC and TB+MALA beat OSIS in most cells of Table 1. We will remove "not consistently outperformed by any baseline" and report that OSIS falls short of those baselines, while remarking that they have much higher computational cost.
  2. In Table 4, TB+MALA had the lower (better) FID, and the comparison uses one prompt and no repeated seeds. The reviewers’ concerns about "lower FID" and "performs similarly" are no longer relevant with our new results (see the updated Table 4) and answers to W1-W2 for Reviewer TfVR.
  3. The outsourced sampling experiments sample in 16- and 64-dimensional latent spaces: we already state in Section 6.4 both the dimension of the sampling problem and that of the decoded image.
  4. We will clarify that OSIS is a fast amortised sampler which does not provide ELBO or ESS, as it is implicit and does not provide a density.
  5. We will add a direct distillation baseline under OSIS's own teacher kernel alone, then fit flow matching to its samples.

Training cost and time comparison

Quality vs cost. We compare OSIS to the baselines in terms of training and sampling cost as measured using number of energy evaluations (Training E), energy gradient evaluations (Training ∇E), and number of training steps (Optimiser updates). The results are presented in Table G1.

Although OSIS requires more training steps, it uses about 6× fewer training gradient evaluations than DIS or PIS (about 30× fewer in the 2000 outer iterations NICE-FMNIST configuration, T=2000) and samples with one network evaluation instead of 128 network evaluations and 128 gradient evaluations.

Table G1: Training and sampling cost on NICE-FMNIST (d = 784). OSIS (T=2000) uses T=2000 outer iterations instead of T=10 000 for the paper config. All other parameters are identical.

MethodOptimiser updatesBatch sizeTraining ∇ETraining EPer-sample costW₂²MMD(×10⁻³)
OSIS (paper)5.0 × 10⁵4,0961.6 × 10⁹1.6 × 10⁹1 NN1743.96 ± 180.58382.13 ± 59.92
OSIS (T=2000)1.0 × 10⁵4,0963.3 × 10⁸3.3 × 10⁸1 NN1477.98 ± 23.51327.10 ± 22.60
FAB (flow only)1.2 × 10⁴1,0003.7 × 10⁸3.7 × 10⁸1 NN1768.75 ± 335.78430.25 ± 72.05
FAB1.2 × 10⁴1,0003.7 × 10⁸3.7 × 10⁸120 NN, 120 ∇E, 120 E1746.47 ± 335.35438.79 ± 71.94
DIS4.0 × 10⁴2,0001.0 × 10¹⁰8.0 × 10⁷128 NN, 128 ∇E5321.68 ± 2656.55421.53 ± 124.32
PIS4.0 × 10⁴2,0001.0 × 10¹⁰8.0 × 10⁷128 NN, 128 ∇E2793.73 ± 19.11290.08 ± 4.09
TB+MALA4.0 × 10⁴2,0001.0 × 10¹⁰2.7 × 10⁷128 NN, 128 ∇E3074.87 ± 23.88430.23 ± 4.81
SubTB+SMC4.0 × 10⁴2,0001.7 × 10¹⁰1.7 × 10⁹128 NN, 128 ∇E, 128 E4035.49 ± 82.91531.04 ± 15.03
SMC––002,805 ∇E, 1,020 E2856.36±15.26629.18±3.24

We also compare the algorithms in terms of training time. We find that OSIS takes less time to train on most targets and gives the best W₂² on the highest-dimensional task, NICE-FMNIST (Table G2). FAB is substantially faster (Table G2), but its one-step version performs worse or comparably on most targets (manuscript’s Tables 1 and 2). Moreover, unlike FAB, OSIS needs no invertible or tractable-density architecture (e.g., a UNet on NICE).

Table G2: Training time on one GH200, in minutes. Baselines (JAX jit) and OSIS (PyTorch torch.compile). Run with 1 seed.

MethodGMM2GMM10Funnel10MW32MW64NICE-MNISTNICE-FMNIST
OSIS48.647.544.548.548.8278318.2
OSIS (shorter training)10.910.110.510.010.248.658.0
FAB5.76.96.86.86.818.623.4
DIS18.318.617.915.516.5271.3319.9
PIS18.218.518.015.316.0270.0321.7
SubTB+SMC128.3129.1140.2131.6129.2553.3651.5
TB+MALA140.0140.0133.5134.6136.7143.7144.0
AuthorsOpenReview ↗

Response to Reviewer oNmA (1/2)

We would like to thank the reviewer for the very detailed review and for raising important questions about the paper. We are pleased to see the reviewer acknowledging the clarity of our paper, as well as the performance of OSIS at scale on NICE targets and in the latent space of a pretrained GAN. Next, we provide our response to the questions raised in the review.

(W1) Main claims stronger than results support

First, we highlight that OSIS performs well on higher-dimensional targets, beating TB+MALA and SubTB+SMC on both NICE targets (mean W2 and MMD). Second, the results need to be considered in light of the training and inference time: OSIS trains and samples faster than TB+MALA and SubTB+SMC, with no energy evaluations at inference and fewer in training. We give a detailed comparison in Tables G1 and G2 of our general response. It highlights that OSIS has an edge when one can afford imperfect samples but needs them generated quickly with no energy evaluation. Please also see the general response.

We also highlight that OSIS's contribution is broader, including fixed-point and Wasserstein-flow analysis, connections to existing methods such as drifting models, and a systematic study of its design space.

(W1-W2) 40GMM failure

We agree on the failure of OSIS on GMM targets. According to Table 11 (long run of MCMC outperforms OSIS on GMM), we expect that FM distillation with MCMC samples outperforms OSIS on GMM. The reason is that OSIS trains using short MCMC chains, which does not allow it to recover modes. We argue that long chains + MCMC distillation, while efficient on GMM, are strongly dependent on which samples are drawn first, while OSIS resamples at each training iteration. This study will be added to the appendices.

We also agree with the reviewer that the theory’s dependence on the LSI constant helps explain the 40GMM failures rather than resolve them. (Indeed, under α\alpha-LSI, the convergence rate depends on α\alpha. Using, for instance, Section 4.1 in Hong-Bin et al. (2021), we can show an α\alpha log Sobolev inequality for the mixture of Gaussians where the upper bound on the constant alpha increases with respect to the dimension.) But we do not necessarily see this as a weakness. It indicates that one could further investigate the theory to improve it hand in hand with the algorithm, with the aim of obtaining convergence results independent of α\alpha in future work.

Empirically, some inductive-bias heuristics can be used to address 40GMM convergence issues. Specifically, to evaluate such a heuristic, we introduce a jump-MALA MCMC kernel: 10 Metropolis–Hastings random-walk jumps before the 20 MALA steps. The jump size is σk=2hτ2k\sigma_k = \sqrt{2h} \sqrt{\tau} 2^k, with kk cycling through 0,…,50,\ldots,5 , so it is set from the step size and temperature and does not require additional tuning. This substantially improves results on GMM40; in 2 dimensions, it reduces W2 from 103.26 ± 41.41 to 8.73 ± 1.55 (best baseline score 10.20 ± 4.11) and MMD from 64.37±13.38 to 10.50 ± 0.20 (best baseline score 9.50 ± 2.66); sample visualisations are available here [samples]. This is not a solution to OSIS in all settings, as MH would struggle more with sampling from modes in higher dimensions, but it shows that OSIS, as a general method, is relevant and that, under well-designed kernels, it can also sample from well-separated distributions.

Hong-Bin et al., “Dimension-free log-Sobolev inequalities for mixture distributions”, Journal of Functional analysis, 2021.

(continued in next comment)

AuthorsOpenReview ↗

Response to Reviewer oNmA (2/2)

(W2-Q3) Novelty concerns

Cao et al. (2026) is concurrent work, and neither it nor Li et al. (2017) has released code, so we cannot compare against exactly what those papers ran. We therefore implemented both methods from their descriptions and ran them on our benchmarks. We found that neither solves the benchmark problems satisfactorily:

  • Li et al.’s adversarial training (AMC) was unstable, and we could not make it converge on 40GMM, Funnel, or ManyWell. We attribute this to the min–max game between generator and critic, which is known to be fragile and prone to mode collapse (Mescheder et al. (2018)). OSIS replaces the critic with a fixed divergence and leverages inner iterations to remove that instability.
  • Cao et al. performs poorly on our targets, which are harder and higher-dimensional than those it was demonstrated on (low-dimensional Gaussian mixtures, banana and Bayesian neural net posteriors). We believe the cause is its drift estimate, built by Monte Carlo or from second-order derivatives of the energy. Its variance and cost grow with dimension. We include this method in Table 1 for completeness.

Moreover, we discuss the connections of our method to prior work in Section 3, and we believe that making these connections is an important strength of the paper, for the following reasons:

  • Each of these prior methods relies on a specific choice of divergence (e.g., KL) or kernel (e.g., SVGD). In contrast, we show that all of them can be viewed as instances of a single unifying algorithm, OSIS, in which the divergence and kernel are configurable design choices. We propose a theoretical study of our idealised algorithm in Wasserstein space and connect it to existing methods such as drifting models. Although we do not compare against each method individually, we run ablations (see Tables 3, 10, 12, and 13) to identify the best design choices.
  • We are the first to systematically evaluate one-step sampling methods on modern sampling benchmarks and scale it to complex targets like NICE and problems like sampling under a pretrained generative model prior, which were not considered in past work on one-step sampling.
  • We are the first to connect these algorithms — OSIS, and by extension the past algorithms that are nearly special cases of OSIS — with recent developments in one-step generative modelling.
  • We give the first known theoretical results showing that one-step algorithms based on targeting local stochastic refinements (MCMC) inherit theoretical properties of those refinements.

We therefore believe that the proposed algorithm, with its customisable choice of divergence and kernel, together with the evidence of its scalability and the accompanying theory, is a meaningful and valuable contribution to the field.

Mescheder et al., “Which Training Methods for GANs do actually Converge?”, ICML 2018.

(W2) OSIS convergence rate

We agree that OSIS converges at best at the Langevin rate. The OSIS theoretical study is in terms of consistency and convergence; it is not an acceleration result. The strength of OSIS is its amortisation.

Indeed, the case of η=1\eta=1 is favourable, and this is one of the reasons why we introduced the inner iterations. By iterating gradient steps after each rollout, we aim to bring the generator distribution closer to its MCMC rollout while reducing the gradient noise. We observe that the balance between MM and η\eta becomes important with the minibatch approximation, where too many inner steps on a single finite batch can overfit it.

Although it is possible to use W2\cal W_2 as a loss function, and we do experiment with this (see Tables 10, 11, and 13), we empirically found that sliced Wasserstein distance represents a computationally efficient alternative. Connections between 2-Wasserstein and 2-sliced Wasserstein distance can be found in Cozzi and Santambrogio (2025), and Bonnotte (2013), Theorem 5.1.5. A precise study of convergence under sliced Wasserstein lies outside the scope of this paper and would be a substantial theoretical extension that we leave for future work.

Bonnotte, “Unidimensional and Evolution Methods for Optimal Transportation”, 2013.

Cozzi and Santambrogio, “Long-Time Asymptotics of the Sliced-Wasserstein Flow”, SIAM Journal on Imaging Sciences, 2025.

(Q1) Total training cost

Please see Table G1 in the general response.

(Q2) Clarification of the table configurations

The reviewer's reading is correct. In Table 3, the MALA row repeats the Table 1 (M = 50) values in error. The ablation in Table 3 was run for M = 1 with OT loss on GMM40 and SWD on MW32; the MALA metrics have been fixed in Table 3. They do not change substantially and the interpretation is not affected. Thank you for catching this.

Finally, Section D.1, referring to Tables 1 and 2, reports that we use T=10 000 training steps with M=50 inner iterations. All other tables use M=1, in particular, Table 13 uses T=50 000 training iterations for M=1.

Hi Authors,

Thank you for the response. I think most of my concerns have been well addressed. I will raise my score to weak acceptance.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.