acceptodds
Under review as a conference paper at ICLR 2027

Exact Global MCMC with Denoising Diffusion

Abstract

This work shows that diffusion models learned with the standard denoising loss provide exact and effective global MCMC proposals for complex high-dimensional target densities. A forward noising process followed by its exact reversal is a Markov chain that leaves the target invariant, and a denoising diffusion model is a learned approximation of that reversal. A Metropolis–Hastings (MH) step whose acceptance ratio includes the forward and reverse path densities of a discrete-time SDE makes the chain exact for any denoiser, with an acceptance rate that measures how closely the denoiser approximates the ideal reversal. We train denoising diffusion models on locally convergent samples from the Metropolis-adjusted Langevin algorithm (MALA), whose weighting across modes need not match the target since the MH step absorbs the mismatch, and call the composition of the resulting global path sampler with a local MALA sampler Denoising Diffusion Monte Carlo (DDMC). Across Gaussian mixtures, particle systems, and a 550-dimensional Bayesian neural network posterior, DDMC matches ground-truth energy distributions at the measured estimator floor and recovers correct mode weights. On the LJ-38 double-funnel potential below its melting temperature, DDMC obtains the equilibrium funnel occupancy from a mis-weighted training dataset, agrees with parallel tempering, and attains an order of magnitude lower variance at equal total compute, dataset generation and denoiser training included.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
MCMC, diffusion, neural sampler, multimodal sampler, unnormalized density

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_T3m5OpenReview ↗
Rating
3(+2)
Confidence
4
Informativeness
2(+1)
Presentation
3
Soundness
2(+1)
Read review
Paper summary

The paper proposed DDMC:Denoising Diffusion Monte Carlo. It combines MCMC with diffuson models and run MCMC on the path space of the learned diffusion. The rejection is calculated by the forward-backward path weight. The method was evaluated on a large range of tasks, showing its effectiveness.

Critical strengths

The experiment in this paper is comprehensive: GMM, particles and BNNs. This show their proposed method can be applied to a variety of targets. The writing of the paper is also easy to follow and well-structured.

The proposed i-SIR parallel proposals and path tuning might be potential in inspiring future works.

Critical weaknesses
  1. The novelty is limited and missing the most relevant related work. The idea of applying MCMC to path is also standard. For example [1] (eq, 6 and the line below eq 6: Reweighing/resampling methods utilized in this context are e.g. Importance Sampling or Neural MCMC].); [2]. The main contribution of this paper is to apply this to diffusion models. Another highly related work is accelerated parallel tempering (APT) [3]. APT (APT-diff) propose to run a MCMC chain on ladder of paths, while DDMC proposed to run a MCMC chain on the entire paths: DDMC is just the 1 ladder (most vanilla) version of APT in this sense. However, the paper did not mention or compare with them.

[1] Stochastic Normalizing Flows, NeurIPS.

[2] Advanced MCMC Methods for Sampling on Diffusion Pathspace.

[3] Accelerated Parallel Tempering via Neural Transports. ICLR.

  1. One concern I have is the scalability of the proposed method: running MCMC on path is known to be hard. This is because the path largely augment the space and can have much higher variance (hence higher rejec rate). Therefore, how does the method scale with dimension is not super clear to me. When clear evidence supporting my concern is LJ13 to LJ55, the proposal acceptance largely decrease. (also see my question)
Questions for authors
  1. Can you try to see how the rejection increase as the dimension scales?

  2. Standard MCMC can easily tune step size to increase acceptance. What is the analog in this method?

7 comments

We thank the reviewer for directing us towards important prior work.

W1, related work. Our revision now cites and discusses SNF, APT and Beskos et al. (2013). The DDMC acceptance ratio is a realization of the forward/backward path ratio of NCMC, SNF, SPS, and APT that we formulate as an MH correction of the proximal sampler. This proximal sampler correction was introduced concurrently by MAD-Path. We discuss similarities and differences with APT below.
DDMC vs. One-Rung APT. DDMC is closely related but not equivalent to one-rung Diff-APT. DDMC is a Metropolis-within-Gibbs framework that corrects bias in a learned diffusion used as a proximal sampler, while one-rung Diff-APT is a transport between two distributions which evolve independently between swaps. In particular, the noisy state of DDMC is a forward process sample conditioned on a paired target sample at each step, which is not true of one-rung Diff-APT. The two coincide as σmax→∞\sigma_\textrm{max} \rightarrow \infty in which case both become independence samplers. We observe that a finite σmax\sigma_\textrm{max} can provide better mixing than increasingly large σmax\sigma_\textrm{max} (Appendix D.2, Table 7). One-rung Diff-APT must directly model the noisy state marginal at finite σmax\sigma_\textrm{max} using an EBM, which is a limitation not shared by DDMC.
Motivation of DDMC vs. APT. APT contains the single-transport correction as a special case it neither investigates nor recommends. Its premise is that a learned transport is too weak to be accepted on its own, so the ladder exists to split the transport into short hops that PT can tolerate, and its round-trip analysis treats the per-rung rejection as the quantity to keep small. Every APT experiment uses N≥4N \geq 4 rungs and nothing in the paper indicates that one transport level could be viable. DDMC starts from the opposite premise, that a standard denoiser trained on local MCMC data can be accurate enough to serve as a direct global proposal, and the evidence of the paper is that this holds in a variety of cases. DDMC is not a special case that APT covers but a design in the regime APT is built to avoid, and showing that a direct global proposal can work is our contribution.
Score vs. EBM Diffusion. The rungs of Diff-APT are the noised marginals of the target, and PT needs each rung as an unnormalized density for its local moves and swap weights. Diff-APT must therefore parametrize the noised marginals with an energy-based model, take its gradient as the score of the transport, and train it by score matching on its own approximate samples. The swap acceptance then depends on the learned rung densities being right as well as on the transport, which is a source of error that the transport alone does not have, and a score that is constrained to be the gradient of an energy is more restrictive to train and to scale than an unconstrained denoiser. DDMC never evaluates a noised marginal density. The target enters only at the path endpoints, the intermediate distributions belong to the model, and the model can be trained by standard denoising regression instead of EBM diffusion.
Comparison with APT. To our knowledge no APT code or models have been released, so we reproduced its d=100d = 100 scaling study with exact analytic densities and score, an upper bound on any trained Diff-APT, and compared wall-clock per effective sample on one GPU at each method's saturating batch (Appendix F.3, Table 12). DDMC trained on mis-weighted MALA data produces an effective sample 2.4x faster than analytic Diff-APT at its best paper setting K=5K = 5, and 4.3x faster than a hypothetical Diff-APT whose trained model reproduces the analytic one. This adds important context about the viability of direct global proposals beyond what is presented in the APT paper.

W2 and Q1, acceptance against dimension. Please refer to the newly added Appendix H.1 Table 15 and the general response about scaling with dimension. Global proposals for high-dimensional targets are an active research area, and the requirement of a model that is accurate per coordinate applies to the whole class rather than to DDMC alone. Within this class DDMC scales well: it is an exact sampler at the scale of LJ-55, where recent neural samplers struggle to match the target energy spectrum (Table 2, Appendix F.2), and it attains useful acceptance at d=550d = 550 on the BNN.

Q2, the step-size analog. σmax\sigma_\textrm{max} is the proximal sampler's step size: it sets how far a proposal can move, and larger values cross between regions but attain lower acceptance. The same applies to DDMC, which is a correction for a learned proximal sampler. Appendix D.2 (Table 7) scans it on LJ-38 and finds the accepted cross-funnel rate peaks at 0.7, 2.1x the rate at the production value of 2.0, so the main-text LJ-38 results are a lower bound for the same model. This analysis was done after sampling was completed and is not part of the cost comparison.

Sorry for the confusion about my earlier point regarding APT. I agree that APT and DDMC are different in their respective settings, and that DDMC is more flexible.

I also realized that I had misunderstood part of the proximal step. Thank you for answering my questions and clarifying this. I reevaluated the paper and changed my score for now.

I have one follow-up question: How is σ\sigma tuned? For larger σ\sigma, the acceptance rate tends to be low, while smaller σ\sigma values make the proposals less exploratory.

DiGS has a similar issue. I remember that they use two different values of σ\sigma (at least in some of their experiments). When the paper first came out, I checked their acceptance rates and found that almost all accepted moves occurred at the smaller σ\sigma. This may not be a major problem for DiGS (although the overall acceptance rate could be somewhat misleading), since their computational cost is not dominated by diffusion-network evaluations. Here, however, rejected proposals here could waste many expensive network calls.

I also checked the PT baseline settings in G3 and think the baseline could be improved in two ways:

  1. PT should attempt swaps at every iteration rather than waiting for 50 MALA steps between swaps. This would not require additional energy evaluations.
  2. Instead of using a geometric temperature ladder, following Syed et al., PT could use the first few rounds to adapt the ladder so that the swap rejection rates are approximately equal across adjacent temperature pairs.

Thank you for your response and for taking the time to re-evaluate our work. We are glad the clarification of the proximal step was helpful, and we are grateful to you for bringing SNF, Beskos et al. (2013), and APT to our attention. Situating our work in the appropriate context with prior work is very important and these references are necessary for accurately presenting our work to future readers.

Tuning of σmax\sigma_{\textrm{max}}. In this work our goal is to primarily study values of σmax\sigma_{\textrm{max}} that facilitate global proposals. Although smaller values have higher acceptance and require fewer diffusion steps, they are unlikely to provide jumps between regions separated by large barriers, which are a key limitation of existing MCMC techniques. We set σmax\sigma_{\textrm{max}} using heuristics similar to [a], [b], so that the VE-SDE noise spans across the state space. While this level of σmax\sigma_{\textrm{max}} does enable cross-region jumps in our experiments, we noticed that there is still non-negligible dependence on the clean state and that the method is noticeably more likely to propose jumps to the same basin than we would observe for an independence sampler (which can be approached in our framework by using much larger σmax\sigma_{\textrm{max}}). The most useful way to tune σmax\sigma_{\textrm{max}} given a sampling objective is to perform a sweep of different values and see which one provides the most efficient way to measure the quantity of interest. For example, in our post-run analysis of the LJ-38 medium temperature experiment (see Appendix D.2), we found that setting a smaller value of σmax\sigma_{\textrm{max}} than we used in the original run (going from 2.0 to 0.7) might lead to a boost in efficiency of occupancy estimation. In the actual LJ-38 experiment, we decided to use the value 2.0 that is used by iDEM after we had observed round-trips between the funnels at this value in early development experiments before our final training and sampling run.

On the cost of rejected proposals. We see the cost of rejected DDMC proposals as part of the mixing problem rather than as waste. Local samplers also spend most (sometimes all) of their compute on moves that do not change the metastable region, and in PT an entire ladder advances together to gradually produce each cold-rung state whose funnel identity is independent of its past, so what matters in every case is the cost per independent crossing. In DDMC, rejected global proposals are part of the independent crossing cost. In practice we recommend a global neural proposal only in situations like LJ-38 below melting, where local samplers like MALA fail to mix and PT is known to be slow.

PT Baseline. Thank you for your suggestions on improving our PT baseline. Making a fair wall-clock comparisons for MCMC and deep learning methods is difficult, and we devoted considerable effort to tuning the PT baseline. Our ladder was chosen from a screen that varied the number of rungs, the hottest temperature, the rung placement, and the number of local steps per swap round. The compute needed for a converged occupancy estimate varied by less than a factor of two across the configurations we tried, which suggests the baseline is in the right ballpark. Nonetheless, we expect there is still room for improvement in the PT tuning. Attempting swaps after every local step is a helpful suggestion and worth trying in future PT baselines. We did not use the adaptive spacing of Syed et al., but our geometric ladder already gives approximately uniform swap acceptance across adjacent pairs (Appendix G.3).

We hope this addresses your remaining questions, and we thank you again for the careful reading.

[a] Song and Ermon. Generative Modeling by Estimating Gradients of the Data Distribution. NeurIPS 2019.
[b] Song and Ermon. Improved Techniques for Training Score-Based Generative Models. NeurIPS 2020.

Reviewer_kH7ROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Denoising Diffusion Monte Carlo (DDMC) to improve sampling from high-dimensional multimodal distributions, where conventional MCMC methods often struggle to move between modes. The method uses a denoising diffusion model to generate global proposals and applies a Metropolis–Hastings (MH) correction over the full diffusion trajectory to ensure that the resulting Markov chain has the desired stationary distribution. As a result, inaccuracies in the learned reverse process affect the acceptance rate without introducing asymptotic sampling bias. The diffusion model is trained using locally equilibrated samples generated by MALA, and the MH correction allows the method to accommodate differences between the mode weights in the training data and those of the target distribution. DDMC also combines global diffusion proposals with local MALA updates, together with reverse variance calibration and iterative sampling importance resampling (i-SIR). The method is tested on Gaussian mixtures, particle systems, and a 550-dimensional Bayesian neural network posterior. The reported results show improved sampling performance on several benchmarks. Notably, for the low-temperature LJ-38 cluster with its double-funnel energy landscape, DDMC obtains structural occupancy estimates close to those from parallel tempering while reducing sampling variance by approximately one order of magnitude under comparable computational budgets.

Critical strengths
  1. The path‑space Metropolis‑Hastings correction guarantees exact MCMC sampling despite fitting and discretization errors in the diffusion denoiser; model imperfections only degrade acceptance rates without introducing statistical bias.

  2. This approach does not require training datasets to preserve true mode weights of the target distribution and only demands coverage of high‑probability regions. On the challenging LJ‑38 double‑funnel potential, it achieves roughly an order‑of‑magnitude lower estimation variance than parallel tempering under equal computational cost.

Critical weaknesses
  1. Methodological novelty and overlap with existing work. The main methodological contribution of DDMC is not clearly distinguished from MAD-Path and SPS. All three methods use the joint densities of forward and reverse diffusion trajectories to construct a path-space Metropolis-Hastings acceptance ratio, allowing the target distribution to be preserved even with an imperfect learned reverse model. The difference appears to lie mainly in how the reverse model is trained, with DDMC using denoising regression rather than variational score-based training, rather than in the sampling mechanism itself. In addition, the combination of global proposals and local MALA refinement has already been explored in Ex²MCMC. The authors should clarify what is fundamentally new in DDMC compared with these methods, beyond the choice of training objective and the integration of existing sampling techniques.
  2. Computational efficiency and scalability. The improvements reported on LJ-38 are substantial, but the current experiments do not fully establish the computational advantages of DDMC. In particular, it is unclear how the method performs as the problem dimension increases or when the available computational budget and number of parallel proposals change. The cost of training the diffusion model, running long reverse diffusion trajectories, and generating multiple proposals should also be considered when comparing sampling efficiency. End-to-end runtime comparisons under matched computational budgets, preferably including stronger baselines, would provide a clearer picture of the actual efficiency gains. Without these results, it is difficult to judge whether the improvements observed on LJ-38 extend to larger and more challenging sampling problems.
Questions for authors

Dependence on prior knowledge of target modes. In the low-temperature LJ-38 experiments, the training datasets are constructed from a pre-existing database of local minima obtained through basin hopping, which also contains the known global minimum. This raises the question of how much the reported sampling efficiency depends on prior identification of the relevant energy minima. Is the computational cost of building this database and locating the important energy funnels included in the reported budget? Without accounting for this cost, it is difficult to assess whether DDMC can efficiently discover previously unexplored modes or primarily benefits from prior knowledge of the target distribution.

1 comment

We thank the reviewer for the detailed summary and insightful comments.

W1, relation to MAD-Path, SPS and Ex$^2$MCMC. Please refer to the general response about contribution and relation to prior work. Ex$^2$ MCMC takes the global proposal as given, a fixed distribution or an adaptively trained flow, and neither has been shown to attain viable acceptance at the dimensions and geometries we study. MAD-Path lacks an effective way to train the score model (Appendix F.2). The motivation for DDMC is to align the requirements of global proposals, which need a high-quality model, with a standard diffusion learning method known to scale well with data, model size, and compute. DDMC can be an exact global sampler at a scale where recent biased neural samplers do not match the target energy spectrum (LJ-55) and for a target with several hundred dimensions (BNN).

W2, efficiency and scaling. We make several revisions to address this.
Dimension. Please refer to the newly added Appendix H.1 Table 15 and the general response regarding dimension. The requirement of increasingly strict model accuracy as target dimension increases is shared among independence-type global samplers. Among possible frameworks for meeting the high accuracy requirements of global samplers, we argue the standard diffusion training framework of DDMC is a promising and natural direction.
Parallel proposals and ladder length. Please refer to the newly added Appendix H.2, Table 16: on LJ-55, K from 2 to 128 raises acceptance from 0.028 to 0.42 and mixing saturates near 4 rounds from K = 64, while effective samples per GPU-hour on one GPU peak at K ≤\le 8. The i-SIR pool can increase efficiency when parallel capacity is available.
End-to-end budgets. The LJ-38 cost comparison already includes DDMC dataset generation and training (Table 13), and Appendix G.4 plots precision against cumulative compute at matched budgets. To our knowledge PT remains the gold-standard baseline for LJ-38 below melting, and on LJ-55 we compare against a representative selection of strong baseline neural samplers (Table 2, Appendix F.2).

Q, prior knowledge of the modes. Both LJ-38 methods received the same prior knowledge: PT is tuned using chains initialized from separate funnels to optimize cross-funnel movement, and PT's low-temperature ensembles start from the same basin-hopping minima as the DDMC dataset. Basin hopping for LJ-38 costs about 30 minutes on CPU, is shared by both methods and is not charged (see general response about basin hopping). The LJ-38 experiment measures reweighting given coverage, and we do not claim discovery of unrepresented regions (see general response regarding discovery).

Reviewer_4iuPOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces Denoising Diffusion Monte Carlo (DDMC), combining global proposals from a standard denoising diffusion model with path-space Metropolis–Hastings correction and local MALA updates. Auxiliary diffusion paths make the acceptance ratio tractable without evaluating the diffusion model’s endpoint density. Experiments demonstrate correction of inaccurate training-set mode weights, useful acceptance in high dimensions, and improved cost-to-precision on the low-temperature LJ-38 system.

Critical strengths
  1. A clear separation between proposal quality and correctness. The augmented-space construction preserves the target distribution while allowing flexible denoisers and discretizations.

  2. A convincing main experiment. The LJ-38 study evaluates a meaningful structural observable across 128 independent replicas and includes dataset generation and training in the reported compute.

  3. Useful practical methodology. Reverse-variance calibration, acceptance diagnostics, and batched i-SIR proposals make the construction more usable.

  4. Transparent discussion of limitations. The paper acknowledges missing-region risks, long-lived trapping, and settings where MALA remains competitive.

Critical weaknesses
  1. The strongest result depends on prior coverage of the important metastable regions.

    Low-temperature LJ-38 training data are seeded from a basin-hopping database, including the known global minimum, because random initialization almost never reaches that funnel. This is transparently disclosed, but it limits the interpretation: the experiment demonstrates efficient reweighting and transitions between represented funnels, rather than discovery of unknown important regions. Controlled experiments varying basin coverage and training-set imbalance would better characterize practical robustness.

  2. The headline PT comparison requires more complete baseline specifications.

    The reported cost-to-precision improvement is substantial, but the manuscript does not provide enough detail about the LJ-38 PT temperature ladder, number of rungs, local-update schedule, swap frequency, and tuning procedure. These choices strongly affect PT performance. The compute accounting should also explicitly state whether constructing the basin-hopping database, calibration, and hyperparameter searches are included.

  3. The diagnostics do not fully characterize mixing beyond energy and funnel identity.

    Energy-distribution agreement is useful but cannot establish agreement of the full configuration distribution. For example, on LJ-55 the reported sample-space Wasserstein distance remains around 3.88, compared with a reference floor of 2.57, despite matching the energy floor. Reported IACT values are also based on energy. Additional structural observables, observable-specific effective sample sizes, and convergence from deliberately difficult initial states would strengthen the finite-run evidence.

  4. The practical benefit outside the main LJ-38 experiment is less established.

    Several targets can already be sampled effectively by local MALA. The BNN experiment requires a 305M-parameter denoiser, approximately 280 GPU-hours of training, and concurrent sample generation. Its single-chain initialization advantage does not translate to the 512-chain setting, where MALA is faster, and predictive NLL differences are not statistically significant. End-to-end efficiency curves and break-even sample budgets would clarify when this investment is worthwhile.

Questions for authors

See weakness.

3 comments

We thank the reviewer for a careful reading and constructive criticisms.

W1, coverage and imbalance. Please refer to the newly added Appendix C.3 (Table 6), and our general response about discovery. The imbalance appendix shows DDMC can tolerate a few multiples of mis-weighting but will fail for significant mis-weighting. Adaptive retraining can alleviate this issue. For example, regions where a chain dwells longer than expected (such as the basin in Appendix G.5) can become samples for continued training. If the metastable regions are known but their weights are not, models trained on pairs of regions with 50/50 weighting (at most 2x under-representation) estimate the relative weight of each pair, and these pairwise comparisons can be composed to find the regions with the highest mass and the weights among them. We plan to explore this direction in future work.

W2, PT baseline specification. Please refer to the newly added Appendix G.3 which gives the full PT configuration: a geometric ladder to T = 0.40 with 16 and 20 rungs, 50 MALA steps per round with per-rung adapted step sizes, non-reversible even-odd swaps, and a tuning procedure which included significant effort to find the best parameters for PT to cross between funnels.
Basin hopping. Please refer to the general response. Basin hopping is quick and was used both to train DDMC and tune PT.

W3, mixing beyond energy. Thank you for raising this important point. Tables 3 and 4 now report a Structure IACT next to the energy IACT, and Table 1 adds a mode center IACT for the Gaussian mixtures, which measures how quickly chains move between modes. For the particle systems Structure IACT uses the greater of the IACT of the radius of gyration and the maximum pair distance, and for the BNN the greater of the IACT of the two leading principal components of the weight vector. Structural IACT indicates lower sample efficiency compared to energy IACT and is a more reliable indicator of mixing.
The LJ-55 sample distance of 3.88 against a floor of 2.57. The reference floors in Table 2 use sample banks from prior work. We generated independent parallel tempering references for all three particle systems (new Appendix E.4, Table 9) and recalculated the W2W_2 metrics. On LJ-55 the floor is 3.862 ±\pm 0.005 and DDMCK\textrm{DDMC}_K scores 3.864 ±\pm 0.004, and DDMCK\textrm{DDMC}_K is at the floor on both metrics on all three targets.
Difficult initial states. Every sampling chain in the paper starts from a random normal or uniform initialization, which provides a straightforward target-agnostic method to begin a fresh chain. Chains initialized from states with non-negligible target mass that were not learned well or which received very low weighting can become stuck, and solving this problem requires adding missing states to the training data or changing a state's weighting under the training distribution.

W4, benefit outside LJ-38. We agree that the efficiency claim is confined to LJ-38, where neither MALA nor PT is efficient. The other targets test exactness and acceptance of global proposals at a scale where recent biased neural samplers have difficulty matching the energy spectrum (the LJ-55 and BNN targets). For the BNN target, where MALA can freely mix, we do not expect much practical benefit for DDMC, and the two timings in Table 4 are a snapshot of two scenarios rather than a break-even analysis. The new section Appendix G.4 adds precision against cumulative compute for LJ-38. In the medium temperature case PT burn-in and DDMC data/training take about the same time, while in the low temperature case PT burn-in takes a significant portion of the run budget and has not fully converged by the end of the run. DDMC burn-in is rapid and we do not truncate early samples from the DDMC chains when calculating occupancy.

Many thanks for the response. I will keep the positive score.

Thank you for keeping the positive score and for your careful review. Your suggestions, in particular the request to investigate other IACT metrics, led to revisions that made the paper better, and we appreciate your time and effort.

Comments on the paper

3 comments
AuthorsOpenReview ↗

summary of revisions

We sincerely thank all reviewers for their time and careful readings of our paper, which have led to several important additions and revisions. This note lists the major changes in the revision.

Changes in the revision. All added or changed text is blue, and all section, table and appendix numbers in our responses refer to the revised PDF. New appendices: C.3 (dataset imbalance on GMM-256), D.2 (proximal step size σmax\sigma_{\textrm{max}} on LJ-38), E.4 (independent reference floors for the particle targets), F.3 (comparison with APT), G.3 (parallel tempering baseline), G.4 (precision against compute on LJ-38), H.1 (acceptance and dimension), H.2 (pool size and ladder length on LJ-55). Tables 3 and 4 add a structural IACT next to the energy IACT. One experimental configuration was changed: σmax⁡\sigma_{\max} for the 2D MoG-40 target is 2.02.0 in place of 0.380.38, since the smaller value proposes moves to nearby modes only, and the larger noising level performs much better on the newly added mode center IACT metric with only a moderate decrease in acceptance.

AuthorsOpenReview ↗

general response to reviewers

We sincerely thank all reviewers. Four questions were raised by more than one reviewer, and we answer them here.

Acceptance against dimension. DDMC belongs to the class of global proposals with properties similar to an independence sampler (the σmax→∞\sigma_\textrm{max} \rightarrow \infty version of DDMC), in that a close match between proposal and target across distant regions is needed for the sampler to be viable. We expect that a similar requirement applies to global DDMC proposals which only have a weak correlation with the current state. This requirement holds across independence-like global samplers, not only for DDMC. For an independence proposal the log acceptance ratio has stationary mean E[log⁡r]=−KL(pθ∥q)−KL(q∥pθ)\mathbb{E}[\log r] = -\mathrm{KL}(p_\theta \| q) - \mathrm{KL}(q \| p_\theta), and writing ε=−E[log⁡r]/d≥0\varepsilon = -\mathbb{E}[\log r] / d \geq 0 for the divergence per coordinate, the acceptance is at least e−εde^{-\varepsilon d} by Jensen's inequality. A per-coordinate divergence of order 1/d1/d is therefore enough to hold acceptance fixed as dd grows, and our Lennard-Jones results show that acceptance does fall when ε\varepsilon stays constant and dd grows. The analogous statement for importance sampling is that a sample size of order eKL(q∥pθ)e^{\mathrm{KL}(q \| p_\theta)} suffices (Chatterjee and Diaconis, 2018). Appendix H.1 (Table 15) reports −log⁡(accept)/d-\log(\mathrm{accept})/d on every target, which is a lower bound on ε\varepsilon by Jensen and is the quantity that sets efficiency. Across LJ-13, LJ-38 and LJ-55, which share one architecture and training recipe, this cost stays at 0.020.02 to 0.030.03 nats per coordinate from d=39d = 39 to 165165, so the per-coordinate error of the denoiser did not grow with system size, and acceptance decreased as expected. The motivation for DDMC is to align the requirements of global proposals, which need a high-quality model, with a standard diffusion learning task known to scale well, so that better training within the diffusion framework is a direct lever on global move acceptance.

Contribution and relation to prior work. DDMC makes the proximal sampler exact for a learned discrete reverse process by an MH correction on the forward and reverse paths. This proximal sampler correction is a recent formulation proposed concurrently by MAD-Path. It is a specific realization of the forward/backward path ratio that appears in several prior works such as NCMC, SNF, SPS and APT. The local-global composition with MALA follows Ex$^2$MCMC. The revision states this lineage in Sections 1, 3 and 4. Our main contribution is the demonstration that a model trained by standard denoising regression on locally converged, mis-weighted MALA samples is accurate enough to be accepted as a direct global proposal up to d=550d = 550, with exactness for any denoiser. This gives an exact global sampler at a scale where recent neural samplers do not match the target energy spectrum on LJ-55 (Table 2, Appendix F.2) and where the DiGS global step is not accepted on the BNN (Appendix F.1), and it recovers funnel occupancy on LJ-38 below melting in agreement with PT with much less variance at equal compute (Section 5.4, Appendices G.3 and G.4).

Discovery of metastable regions. We do not claim DDMC can effectively discover high-mass metastable regions that are absent from the training data. The low-temperature LJ-38 experiment measures reweighting given coverage: the dataset is split 50/50 across the two funnels and DDMC recovers π(Ih)≈0.007\pi(I_h) \approx 0.007, a reweighting by two orders of magnitude, in agreement with parallel tempering. The log of the occupancy ratio of the two funnels is their free energy difference in units of kBTk_B T, which sets the relative stability of the two structures, and such differences between known metastable states, for example protein conformations, are a central quantity in molecular simulation. In our framework, discovery belongs to the construction of the training dataset and not to the sampler. It often requires domain knowledge, and it is a problem shared by every neural sampler trained on target samples. Appendix C.3 (Table 6) measures how far the training weights can depart from the target before the sampler fails: a 6.7×6.7\times under-representation of a mode is recovered at about 8×8\times slower mixing, while a 67×67\times under-representation is not. Adaptive retraining from the chain's own samples, and dataset design more broadly, are important research areas for neural samplers.

Basin hopping cost. Basin hopping on LJ-38 takes about 3030 minutes on CPU, and it informs both DDMC training and PT tuning. The DDMC training dataset is produced from its minima. PT was tuned with ensembles seeded in opposite funnels discovered via basin hopping so that its ladder is selected to maximize the mixing speed between the funnels (Appendix G.3). The cost of basin hopping is small, shared, and is not charged to either method (Appendix G.4).

Area_Chair_AXQmOpenReview ↗

Author-Reviewer Discussion ends in 1 day, 12 hours

Dear authors and reviewers,

Thank you to the authors for their detailed responses, the new appendices and the revised manuscript. The discussion period closes in about a day and a half, so this is the last chance for a direct exchange before the reviewer-AC discussion starts on Nov 19.

Reviewer 4iuP: I haven’t yet seen a reply from you. Please read the authors’ response and the revision, especially the new appendices which address your points. Let us know whether your concerns have been addressed, and please update your recommendation if your assessment has changed.

Reviewer kH7R: thank you for your reply. Since you are keeping your score, could you briefly say which concerns remain open ? That would let the authors respond and help me weigh them in the next phase.

Reviewer T3m5: thank you for engaging and for re-evaluating the paper. The authors replied on Nov 17 to your follow-up. Please let us know whether this settles your remaining points.

Authors: please answer any new questions in the time left, focusing on the points most likely to affect the decision.

Thank you all for keeping the discussion constructive and timely.

Your AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.