WorksMathematics
Amortizing Tempered Sampling with Corrected Self-Training
Abstract
Learning a neural sampler can reduce the cost of repeated sampling from an un-normalized distribution, but obtaining good training data remains difficult. We develop a corrected self-training scheme across a sequence of tempered targets, making use of importance sampling and Markov chain Monte Carlo to correct the samples and replay buffers to store them. We introduce Sequential Score Sampler (SSS) as a particular case. In SSS, the proposal is a score-based diffusion trained by regression on energy gradients; and the samples are corrected by importance weights on the path space. After training, samples can be collected directly from the model, from its importance-corrected proposals, or from the final-temperature buffer. The method amortizes repeated sampling by encoding the global structure of the target in the model. SSS achieves competitive sampling accuracy across the benchmarks and, on several targets, uses shorter runtimes with fewer energy evaluations than competing samplers.
Introduction
Many inference and simulation problems require repeated sampling from a distribution whose density is known only up to a normalizing constant. Multimodality makes this difficult: local Markov chain Monte Carlo (MCMC) may explore a basin thoroughly while rarely moving between basins [44]. Annealed importance sampling and sequential Monte Carlo (SMC) address this difficulty through intermediate distributions [28, 9, 13, 11]. However, obtaining further accurate estimates still requires additional sampling computation. Can some of that computation instead be invested in learning a reusable proposal?
Learned samplers pursue this amortization [46, 39, 27, 1]. Once trained, a generative model can generate new proposal samples without continuing a long local Markov chain. The difficulty is how to train it without representative target samples. One approach is to generate samples from the current model, correct for proposal mismatch, and use the resulting data to update it. Importance weighting supplies this correction by adjusting the samples’ contributions without requiring the target’s normalizing constant.
We develop corrected self-training along a sequence of tempered targets and instantiate it with Sequential Score Sampler (SSS). Rather than fit a separate proposal at each temperature, we train one model that takes the temperature as an input. Initial sample populations can come from a suitable sampler or existing problem-specific simulations. During training, the model generates proposals at each intermediate temperature. We resample them according to their importance weights and apply MCMC steps that preserve the corresponding target. Reweighting to the next temperature then provides data for its buffer. These buffers retain weighted samples from recent iterations, allowing to reuse costly sampling results over multiple optimization steps instead of training only on the latest batch. Each outer iteration updates the model using data from all temperatures.
In SMC with local Markov transitions, reweighting and resampling cannot create configurations outside the current population; reaching an unrepresented basin therefore depends on rare transitions. A learned global proposal can instead reuse information encoded from earlier samples and across temperatures [14], without starting from a surviving particle. Problem-specific architecture can support this reuse. For interacting-particle systems, for example, an equivariant network shares learned geometric relations across rotations and permutations of identical particles (Satorras et al., [34]). Such structure reduces the need to learn equivalent configurations separately and can support proposals beyond those stored in a finite population.
Section 2 presents the general training scheme, independently of the proposal family and initialization sampler. Section 3 gives the score-based implementation. Section 4 asks how much accuracy is retained by the learned proposal and how much is gained by correcting its samples or drawing from the final buffer. We compare all three outputs with Temperature-Annealed Boltzmann Generators (TA-BG; Schopmans and Friederich, [35]), Progressive Inference-Time Annealing (PITA; Akhound-Sadegh et al., [2]), and Adjoint Sampling (AS; Havens et al., [20]).
Corrected self-training across temperatures
Targets and initialization
Let be a normalized reference density from which sampling is easy and whose support covers the target. We write the target as and define the intermediate unnormalized and normalized densities
Here , so and . We call the ‘temperature’ parameter. Choose a ladder and obtain initial weighted populations approximating the targets at its levels. These choices may use problem knowledge, existing simulation data, or a preliminary sampler such as SMC or parallel tempering. Adjacent targets should overlap sufficiently for incremental importance reweighting. Our adaptive-SMC initialization is one particular implementation of this step.
Let denote the distribution generated by the learned proposal when given temperature . The same parameters are shared across all levels; in a neural sampler, they are the network weights. We write for its density.
For each target , we maintain a weighted training dataset , called a replay buffer, initially filled with the corresponding population. Retaining data across refreshes is useful when generating corrected samples is expensive relative to a model update: each sample can be reused in several optimization steps, and training need not rely on a single, potentially noisy batch. Pooling recent batches also retains configurations from earlier iterations while the proposal changes.
Training alternates between refreshing these datasets with corrected proposals and fitting the shared model to data from all temperatures. A refresh at supplies data for . Figure 1 summarizes the data flow, and Algorithm 1 specifies the update order.

Figure 1. Corrected self-training across temperatures.
Generating corrected training data
For each selected level , we refresh using corrected proposals from the preceding, typically easier target . Well-initialized buffers may be kept fixed; our choice of refreshed levels is specified in Section 4.1.
Generating weighted proposals at beta k
At fixed model parameters , generate proposal samples and nonnegative importance weights targeting . We require the importance-sampling consistency property: for test functions ,
When the proposal density is available, the weight is simply . Otherwise, importance sampling can be performed on an extended space that includes the sampler’s auxiliary variables, provided the extended target has marginal . § 3.2 gives such a construction for diffusion paths.
Resampling and applying MCMC
We resample the proposals according to their importance weights and apply -invariant MCMC updates. The weights adjust the proposals’ contributions to the target approximation. The Markov moves can diversify duplicated particles and explore nearby configurations while preserving the target distribution.
Reweighting and updating the buffer at beta k+1
To turn the resulting population into training data for the next target, assign each sample the incremental importance weight
Unlike the proposal-correction weights , these weights account only for the change of target from to . We insert this weighted batch into . In our implementation, each buffer has fixed capacity and uses first-in, first-out (FIFO) replacement: new batches gradually replace the oldest entries. This retains data across several refreshes without keeping all past populations indefinitely. Before insertion, we normalize weights within each incoming batch to have mean one, giving equal-sized batches equal total weight. Appendix § 7.1.§ 2 gives the details.
Learning one model from all temperatures
After refreshing the selected buffers, we draw training samples from every with probabilities proportional to their stored weights. Each temperature contributes the same number of samples to a training minibatch. The fitting objective depends on the proposal family: a normalizing flow can be trained by weighted maximum likelihood, whereas SSS uses the score-regression loss in § 3.1. Gradients are taken only through this fitting objective, not through sample generation, importance weights, resampling, or MCMC. Sharing parameters allows features learned at one temperature to influence proposals at others, while retaining training data from the entire ladder throughout optimization.
Final sample collection
After training, the construction supports three ways to collect samples at the target temperature, evaluated in § 4 as SSS (raw), SSS (corrected), and SSS (buffer).
Raw model samples. Draw directly from , without importance correction or MCMC. The sampling cost is that of the proposal itself, but its distribution may differ from the target.
Importance-corrected samples. Generate proposals at together with their importance weights. These can be used in weighted estimates or to resample the proposals; we evaluate the latter option. The weighted
Algorithm 1 Corrected self-training across temperatures
Input: Reference , energy , ladder , initialized weighted buffers
Initialize the shared model parameters .
for each outer iteration do
for each selected level in parallel do
Generate proposals and importance weights at using
Resample by these weights; apply -invariant MCMC updates.
Compute .
Update with this weighted batch (Section 2.2.3).
Update using weighted minibatches from all buffers.
return The trained proposal, its importance-weight calculation, and .
estimates satisfy (2), but poor proposal-target overlap can concentrate the weights on a few samples, leading to low effective sample size and many duplicates after resampling.
Final-buffer samples. Draw from according to its stored weights. This population has already undergone the correction, MCMC exploration, and temperature-bridging steps during training. Its pool of distinct configurations is limited by the buffer capacity; obtaining new configurations requires further buffer updates.
A score-based implementation
Sequential Score Sampler (SSS) implements the proposal of § 2 using a score-based diffusion. A single network is shared across temperatures, where specifies the tempered target and the diffusion noise level. We describe how the network is trained and how the resulting diffusion generates proposals with computable importance weights.
Learning the relative score
For a given temperature , let . Let denote the density of the Gaussian perturbation for and and where controls the noise level. We train the network to predict the score of this noisy distribution relative to the standard Gaussian distribution :
Thus, is the correction to the known Gaussian score and it can be learned with a standard regression loss. Two standard choices of regression target are:
Denoising score matching (DSM) uses the noise added to (Vincent, [41]; Song et al., [36]), while target score matching (TSM) uses the target score at (Bortoli et al., [6]). Both have conditional expectation given , so either can be used to learn the same function. For a standard Gaussian reference , the TSM target simplifies to by (1) (Appendix § 7.2). The energy gradient can be stored with and reused at different noise levels.
In practice, we train on the weighted replay buffers rather than exact target samples. For each training example, we select uniformly, draw from according to its stored weight, and sample and to construct and the regression target . The training loss is
where controls the contribution of each noise level and temperature.
With exact target samples and sufficient model capacity, either regression target recovers at the population level. In practice, errors in the buffers’ coverage or mode weights can affect the learned score. § 4.1 specifies the loss used in each experiment.
Generating samples and computing weights
We first give the sampler and its weights for a standard Gaussian tempering reference . In this case, the learned relative score defines the reverse-time diffusion
With an accurate relative score and a horizon for which is close to , its terminal distribution approximates . We use an exponential integrator that treats the Gaussian-reference dynamics exactly. Appendix § 7.1.§ 3 gives the numerical update and its implementation. The numerical sampler produces a path starting at , with endpoint . For and , its Gaussian transition density is . Let be the exact Ornstein–Uhlenbeck reference transition. A chain started at and propagated with still has endpoint distribution . Multiplying its path density by therefore gives an extended target with endpoint marginal . The ratio of this unnormalized path density to the proposal path density is
For fixed and integrable , this construction gives , where the first expectation is over proposal paths. Independent simulations therefore satisfy (2). The correction applies to the discrete sampler actually used: it requires neither an exact score nor evaluation of the endpoint density . The weights can be accumulated online without storing the path (Appendix § 7.1.§ 3).
Experiments
Targets, network architectures, and baselines
Targets. We study five targets. Two factorized Gaussian mixtures, GMM16 and GMM50, have the same one-dimensional components but differ in dimension: the number of modes grows from to . Both admit exact reference draws. DW-4 and LJ-13 test whether the sampler reproduces arrangements of interacting particles ([26]). DW-4 has several preferred configurations; LJ-13 also has steep repulsion when particles come close. Finally, we study Alanine Dipeptide at 300 K, where the backbone angles distinguish molecular conformations ([2]). Appendix § 7.3.§ 1 gives the target definitions and reference samples.
Initialization and rejuvenation. In all experiments, one adaptive SMC run selects the temperature ladder and initializes the buffers ([13, 24, 5]). We use Hamiltonian Monte Carlo (HMC), with an empirical preconditioner for the mixtures, DW-4 and LJ-13 and an identity preconditioner in internal coordinates for Alanine Dipeptide, for local rejuvenation; its tuning and the reuse of the SMC-tuned parameters are specified in Appendix § 7.1.§ 1. The reference buffer contains independent -samples. The first bridge weights such samples by , followed by SMC resampling and Markov moves, to initialize . We keep and fixed: the former already contains exact reference draws, and the latter provides an SMC-based training population at the first nontrivial level without relying on a learned proposal. Only are refreshed, using levels in Algorithm 1.
Network architecture. In all experiments, a single network is shared across all temperatures of the ladder, modeling the relative score described in § 3.1. We use a residual multilayer perceptron (MLP) for the Gaussian mixtures, and an -equivariant graph neural network (EGNN) ([34]) for DW-4 and LJ-13. For Alanine Dipeptide we use a diffusion transformer with one token per atom. Appendix § 7.3.§ 2 gives the details.
Training loss. We use target score matching on the Gaussian mixtures, DW-4, and LJ-13. For Alanine Dipeptide we use denoising score matching across the full range of diffusion times . At low noise (), we add a small target score matching term so that the network also learns from the energy gradient of the clean configuration. Appendix § 7.3.§ 5 gives the loss and its settings.
Baselines. We compare SSS with three methods that learn to sample from an energy function. TA-BG and PITA share our use of a temperature ladder. TA-BG first fits a normalizing flow at high temperature. It then reweights its samples to the next temperature and fits the flow to the resampled population [35]. PITA trains a diffusion at each temperature and uses an auxiliary energy model to compute weights along the reverse diffusion, producing training samples for the next temperature [2].
AS shares a different ingredient with SSS: it reuses endpoint gradients to train a diffusion [20]. AS stores endpoints generated by its current sampler. For each endpoint it can draw fresh intermediate states from the base diffusion bridge and regress the control against the stored terminal-cost gradient. In SSS, the endpoints used for training come from buffers at several temperatures. The diffusion supplies its own path weights through (8), and one network learns from the corrected populations across the ladder.
For the two temperature-based baselines, we match the ladders, account for network size, and use the same nominal HMC proposal budget on GMM16, GMM50, and DW-4 (Appendix § 7.3.§ 5). For AS we use an HMC-augmented reimplementation on all five targets, described in the same Appendix section.
Cost and evaluation. We compare each sample set with reference draws using summaries appropriate to the target. For the mixtures, we compare pooled coordinates and energies. For DW-4 and LJ-13, we compare pairwise distances, energies, and whole configurations after alignment. For Alanine Dipeptide, we compare backbone angles and energies, following Akhound-Sadegh et al. [2]. Appendix § 7.3.§ 4 defines each discrepancy. We report energy and gradient evaluations alongside timing measurements in Appendix § 7.3.§ 6.
Main results
Table 1 compares the learned samplers on the Gaussian mixtures and particle systems. The three SSS rows use the same trained model for each target. Across the ten target–metric comparisons in this table, an SSS sample set attains the lowest displayed mean in nine cases, including one tie, and the second-lowest mean in the remaining case.
| Target | Method | Distance- ↓ | Energy- ↓ | Geometric- ↓ |
| GMM16 () | TA-BG | 0.023 ± 0.005 | 0.067 ± 0.043 | — |
| PITA | 0.168 ± 0.030 | 0.984 ± 0.082 | — | |
| AS | 2.221 ± 0.072 | 9.947 ± 0.379 | — | |
| SSS (raw) | 0.130 ± 0.091 | 1.153 ± 0.121 | — | |
| SSS (corrected) | 0.005 ± 0.005 | 0.040 ± 0.004 | — | |
| SSS (buffer) | 0.041 ± 0.009 | <u>0.048 ± 0.016</u> | — | |
| GMM50 () | TA-BG | <u>0.114 ± 0.153</u> | 1.065 ± 1.640 | — |
| PITA | 2.115 ± 0.021 | 240.014 ± 1.016 | — | |
| AS | 2.338 ± 0.060 | 32.965 ± 1.294 | — | |
| SSS (raw) | 0.166 ± 0.034 | 4.304 ± 0.807 | — | |
| SSS (corrected) | 0.048 ± 0.022 | 0.321 ± 0.059 | — | |
| SSS (buffer) | 0.208 ± 0.037 | <u>1.025 ± 0.201</u> | — | |
| DW-4 () | TA-BG | 0.317 ± 0.016 | 24.519 ± 10.463 | 0.789 ± 0.022 |
| PITA | <u>0.150 ± 0.124</u> | 2.664 ± 2.107 | 0.479 ± 0.200 | |
| AS | 0.467 ± 0.090 | 0.914 ± 0.284 | 0.956 ± 0.126 | |
| SSS (raw) | 0.227 ± 0.075 | 0.924 ± 0.313 | 0.523 ± 0.144 | |
| SSS (corrected) | <u>0.192 ± 0.010</u> | <u>0.109 ± 0.034</u> | <u>0.482 ± 0.014</u> | |
| SSS (buffer) | 0.194 ± 0.010 | 0.105 ± 0.014 | 0.479 ± 0.016 | |
| LJ-13 () | TA-BG | 1.21 ± 0.02 | 61.47 ± 0.12 | 4.16 ± 0.01 |
| PITA | 0.04 ± 0.00 | 2.26 ± 0.21 | 1.65 ± 0.00 | |
| AS | 0.899 ± 0.770 | * | 3.179 ± 1.485 | |
| SSS (raw) | 0.041 ± 0.032 | * | <u>1.499 ± 0.043</u> | |
| SSS (corrected) | <u>0.026 ± 0.027</u> | <u>1.405 ± 1.733</u> | 1.626 ± 0.059 | |
| SSS (buffer) | 0.006 ± 0.002 | 0.324 ± 0.063 | 1.465 ± 0.003 |
Table 1. Sampler performance (mean ± standard deviation; lower is better). Bold marks the lowest displayed mean and <u>underlining</u> the second-lowest distinct mean. Equal displayed means share a rank. A — denotes an inapplicable metric and * a mean above . LJ-13 results for TA-BG and PITA are from Akhound-Sadegh et al. [2] and retain the published precision.
Gaussian mixtures. Corrected SSS samples have the lowest mean distance and energy errors on both mixtures. On GMM50, correction reduces Distance- from 0.166 to 0.048 and Energy- from 4.304 to 0.321.
Particle systems. On DW-4, PITA has the lowest mean distance error. The SSS buffer has the lowest energy error and ties PITA on geometric error at the displayed precision (0.479). Corrected SSS ranks second on all three DW-4 metrics. On LJ-13, the buffer has the lowest mean error on all three metrics. The raw SSS samples have a small geometric error, yet their energy error exceeds .
Alanine Dipeptide. Table 2 compares the distribution of backbone angles and the distribution of energies. Raw SSS samples have the lowest mean angular Wasserstein distance, while the buffer ranks first on both energy metrics. Its Energy- is 0.257, followed by corrected SSS at 1.432. Thus the best sample set depends on whether we measure backbone conformations or energies. See Appendix § 7.3.§ 6 for the evaluation budgets.
| Method | Energy- ↓ | Energy- ↓ | T- ↓ |
| TA-BG | 83.457 ± 0.070 | 86.176 ± 0.104 | 0.979 ± 0.012 |
| PITA | 1.530 ± 0.068 | 1.615 ± 0.053 | <u>0.270 ± 0.023</u> |
| AS | * | * | 1.490 ± 0.003 |
| SSS (raw) | 3.765 ± 0.302 | 3.858 ± 0.305 | 0.222 ± 0.058 |
| SSS (corrected) | <u>1.230 ± 1.422</u> | <u>1.432 ± 1.525</u> | 0.757 ± 0.745 |
| SSS (buffer) | 0.212 ± 0.190 | 0.257 ± 0.181 | 0.399 ± 0.207 |
Table 2. Alanine Dipeptide at 300 K. Discrepancies and ranks follow Table 1. Results for PITA and TA-BG are from Akhound-Sadegh et al. [2].
Choice of sample set. A trained model can generate fresh samples whenever needed. The mixture results show how much a subsequent importance correction can improve those draws. The buffer offers a third option when a stored population suffices: on LJ-13 it has the lowest errors across all three metrics, although the number of distinct configurations is limited by its capacity. These differences motivate the next experiments, which examine how correction and local exploration shape the population used for training.
Ablation experiments
We now investigate different aspects of SSS through a set of ablation studies. We run these experiments on DW-4 to examine the roles of importance-correction and MCMC rejuvenation. Specifically, we address two questions. Question 1 (Path-space weight resampling). The diffusion can propose samples in different modes without assigning them the correct relative masses. Does importance resampling improve these masses, or are the retained MCMC updates sufficient?
Question 2 (MCMC rejuvenation). After resampling, MCMC explores nearby configurations and diversify duplicated particles. How does this exploration affect the training data and subsequent learning?
To answer Question 1, we compare SSS with a variant that omits the path-space importance correction: it uses the same architecture, SMC initialization, HMC budget, temperature bridge, and training schedule. Only the resampling by the correction weights is removed when refreshing the buffers. Figure 2 compares the two runs on DW-4. Without path-space correction, both modes of the pairwise-distance distribution remain visible, but their relative masses are distorted. The same distortion appears in the final buffer, and the buffer also acquires a long tail of high-energy configurations. These batches are then reused for training. The normalized path-space ESS, computed from the weights in (8) at , drops sharply when path-space correction is removed. Thus, we conclude that the importance-correction, and not the MCMC updates, is what keeps the training data faithful to the relative masses of the target.

Figure 2. Ablation of the path-space correction on DW-4, final generated samples (orange) and ground truth (blue). Each panel shows the three sample sets of SSS (from left to right): raw samples, importance-corrected samples, and the replay buffer. The top row shows the distribution of inter-particle distances and the bottom row shows the energy distribution.
To answer Question 2, we compare SSS with a variant that omits the HMC steps. Both runs share the architecture, SMC initialization, path-space correction mechanism, temperature bridge, and training schedule. Only the rejuvenation step is removed when refreshing the buffers. Figure 3 compares the two runs at the same intermediate training iteration. Without rejuvenation, the resampled batches consist of a few configurations repeated many times, which appear as sharp spikes in the distance and energy histograms. These spikes are inherited by the training buffer, so the model is trained on a small number of distinct configurations. Consequently, training is slower and less stable. Specifically, at the recorded iteration, the raw samples of the full method already reproduce both distance modes and a plausible energy distribution, whereas those of the variant remain far from the reference and carry a long tail of high-energy configurations. The normalized path-space ESS at reflects the same gap, dropping significantly when rejuvenation is removed. These results indicate that MCMC exploration improves the diversity of the importance-resampled training populations and accelerates learning in this experiment.

Figure 3. Ablation of the HMC rejuvenation on DW-4, evaluated at midway through training (orange) and ground truth (blue). Each panel shows the three sample sets of SSS (from left to right): raw samples, importance-corrected samples, and the replay buffer. The top row shows the distribution of inter-particle distances and the bottom row the energy distribution.
In all of our experiments, we keep the first positive-temperature buffer fixed, choose a particle count for SMC initialization, and reuse its HMC kernel during training. Appendix § 7.5 examines how sensitive the results are to each of these choices.
Related work
Learning proposals with Monte Carlo. Annealed Flow Transport and CRAFT combine learned maps with sequential weighting and MCMC [4, 12]; FAB learns a reusable flow from annealed-importance-sampling computations [27]. Reinforced SMC couples learned proposals and twist functions with SMC, off-policy reinforcement learning, and replay [8]. SCLD likewise combines controlled diffusions with sequential resampling [7]. Other methods also combine amortization and Monte Carlo correction [15, 40, 3]. Our approach trains one sampler with temperature as an input, using supervised learning on repeatedly refreshed, corrected training data from all levels.
Learning across temperatures. PTSD trains diffusion samplers sequentially across temperatures using parallel tempering [37, 18, 21, 16] and MCMC refinement [33]. TA-BG learns a flow through temperature reweighting [35]. PITA combines progressive diffusion training with inference-time annealing to produce lower-temperature data [2]. We share their use of easier targets to train harder ones, an idea also exploited by temperature-conditioned generative models [14, 43, 22], but jointly train a shared model across all temperatures.
Score learning and weighted stochastic samplers. iDEM alternates generation and energy-based score learning [1], while importance-weighted score matching corrects the distribution used in a score-matching objective [42]. Adjoint Sampling derives an endpoint-gradient regression from a stochastic control problem [20]. Its training distribution changes with the current control; our regression approximates a prescribed tempered distribution at each level. Our relative-score target uses known Gaussian identities [17, 41, 32, 6]. Other weighting construction in the literature include Boltzmann generators combining learned equilibrium proposals with reweighting [30, 25, 38], and stochastic normalizing flows computing trajectory weights without marginalizing stochastic transitions [45], building on nonequilibrium path-space reweighting [23, 10, 13, 29].
Discussion and Limitation
Contributions. Our main contribution is a corrected self-training scheme that jointly improves a shared proposal and its training populations across a temperature ladder. The individual ingredients are established, but their combination differs from existing approaches. Unlike progressive training methods such as PITA, which generate training data as temperatures are visited, our scheme repeatedly refreshes corrected populations at multiple temperatures while training one model jointly across the ladder. It also differs from flow-based bootstrapping and learned transport: the model is fitted to weighted populations approximating the tempered targets, rather than learning transport maps between them. At each refresh, proposals are corrected and rejuvenated at one temperature, then reweighted to update the buffer at the next. This allows information collected at one temperature to improve proposals at others, with no need to differentiate through sampling or correction.
Our second contribution is Sequential Score Sampler (SSS), a diffusion implementation of this idea. A single network learns relative scores across temperatures by regression on the corrected populations. The Gaussian transitions provide computable path-space importance weights, allowing proposals to be corrected without evaluating their endpoint density or learning an auxiliary density model. The same training and correction can be used with different proposal families and fitting objectives.
All in all, this leads to a workflow that is simple to explain and to implement and that has demonstrably competitive performance.
Performance. Our experiments evaluate both the benefit of corrected self-training and the performance of the resulting sampler. On GMM50, repeated buffer refreshing reduces the mean distance discrepancy of raw samples from 0.398 with fixed SMC training data to 0.166. Among the compared learned samplers, corrected SSS samples have the lowest mean distance and energy discrepancies on both Gaussian mixtures. On DW-4, the buffer has the lowest mean energy error and ties PITA on geometric error at the displayed precision. On LJ-13, the buffer leads on all three metrics. On Alanine Dipeptide, raw samples give the lowest mean backbone-angle discrepancy, while the buffer has the lowest mean Energy- and Energy-. At a matched energy budget, long-run SMC is competitive on DW-4 and slightly more accurate on LJ-13. This highlights a trade-off: continued SMC can produce more accurate populations at a fixed budget, whereas training also produces a reusable proposal for generating fresh samples.
Limitations. The method depends on adequate coverage of the target distributions by the training buffers. If misses a mode, resampling cannot recover it, and short MCMC runs may fail to discover it. The next buffer can then inherit the gap, which repeated training may reinforce. Even when all modes are represented, concentrated importance weights can lead to low effective sample size and many duplicates after resampling. Local MCMC moves can diversify these copies but may not restore the relative masses of distant modes. Self-normalized importance estimates also have finite-sample bias. Finally, initialization and repeated buffer refreshes incur computational costs, so the value of amortization depends on subsequent reuse of the learned sampler. Appendix § 7.3.§ 7.3.7 discusses the implications for initialization and rejuvenation.
Reproducibility statement
The method is specified by Algorithms 1-2 and Equations (6)-(8), with algorithmic details in Appendix § 7.1 and a justification of the score formula in Appendix § 7.2. Appendix § 7.3 gives the target definitions, network architectures, evaluation metrics, numerical settings, and the description on re-implementation of the baselines. Table 3 reports timings on a single NVIDIA RTX 4500 Ada Generation GPU. The LJ-13 and Alanine Dipeptide results for PITA and TA-BG come from Akhound-Sadegh et al. [2]. Code to reproduce our method is available at https://github.com/VietNH-researching/progressive_is_tempering.git.
| Target | TA-BG | PITA | SSS | AS |
| GMM16 () | 40 | 105 | 6 | 2 |
| GMM50 () | 55 | 190 | 43 | 6 |
| DW-4 () | 80 | 530 | 15 | 13 |
| LJ-13 () | — | — | 552 | 538 |
| Alanine Dipeptide | — | — | 366 | 351 |
Table 3. Wall-clock time (in minutes) for the separately timed runs. All runs were executed on a single NVIDIA RTX 4500 Ada Generation GPU. The LJ-13 and Alanine Dipeptide results of TA-BG and PITA are taken from Akhound-Sadegh et al. [2], so no timing is reported.
Appendix
Algorithmic details
Adaptive SMC initialization
Starting from independent draws from with uniform weights, adaptive SMC chooses each next temperature so that the incremental importance ESS equals a prescribed fraction of the population size. Specifically, at step , reweighting the equally weighted population to a candidate gives the incremental log weights and the normalized ESS
If , we set and the sequence ends. Otherwise, is the root of on . Since is non-increasing in , this root can be found by bisection. We take the left end of the final bisection bracket, where , so each step is slightly conservative. The number of temperatures is therefore determined by , and the target, not set in advance.
The population is then resampled by systematic resampling, applied after a uniformly random permutation of the particles, and moved by MCMC targeting the newly selected distribution . This is the standard adaptive resample-move construction ([13, 24, 19]).
We use preconditioned one-step HMC with full momentum refresh as our MCMC kernel. Let be the symmetrized empirical covariance of the resampled particles at and its Cholesky factor. If is not positive definite, we use instead. Given a step size , one move from is
where . The move is accepted with probability
For fixed , this is ordinary HMC after the linear change of coordinates ; the Metropolis acceptance rule preserves . The step size is chosen by bisection so that the mean acceptance probability of one trial move over the population lies in . Each particle then takes moves (specified in Appendix § 7.3.§ 5). The adapted parameters are then frozen and reused for rejuvenation at during self-training. Each subsequent kernel therefore leaves its target exactly invariant, although invariance alone does not guarantee rapid inter-basin mixing.
The populations obtained at approximate the corresponding targets and initialize with equal weights; is filled with independent draws from . For DW-4 and LJ-13, particles are represented in ambient coordinates and the retained populations are centered before entering the replay buffers. Appendix § 7.3.§ 3 explains the centered target for each system and the corresponding sampling convention.
Replay buffers and per-batch weight normalization
Each buffer has capacity . contains independent draws from ; are initialized with the retained SMC populations. During training, and remain fixed, other buffers are updated every iteration. Specifically, at each active temperature , , we generate new proposal samples per refresh, correct them with importance resampling, rejuvenate with HMC, and insert the resulting weighted samples into . Thus, only have their initial samples removed through first-in, first-out (FIFO) replacement. They are fully replaced after refreshes, subject to the convention for removing partially retained batches.
After resampling according to the proposal importance weights and applying MCMC at , let be the bridge log weights. Store
Here . Each batch has total weight , matching an equally sized batch of initial samples with unit weights, while its relative weights are unchanged. Thus each refresh contributes a self-normalized approximation of the target. Without this normalization, fluctuations in the batch estimate of would also determine its influence relative to other batches.
We use systematic resampling both to correct proposals and to draw training minibatches. The resulting samples can be dependent. Per-batch normalization and additional optimization do not make the finite adaptive buffer an exact sample from its target, and we do not claim convergence of the self-training loop at fixed buffer size.
At each optimization step, we draw a minibatch of samples from each buffer by systematic resampling according to the stored weights, sample fresh diffusion times from the schedule for the target and perturbations , and take one Adam step on (6). We perform such steps between consecutive buffer refreshes. Energy gradients are computed once per stored sample and cached; no gradients pass through the procedure that generates or selects the samples. The value of , the number of outer iterations and the optimizer settings are listed in Appendix § 7.3.§ 5.
Numerical generation and stable log weights
The path weight can be accumulated during sampling without storing the trajectory. Write , , , and . One step is
The proposal and reference transitions have the same covariance, so their Gaussian normalizers cancel. The log-weight increment is simply
The last expression uses the noise already drawn for the step and avoids subtracting two squared norms. Starting from zero, accumulate these increments and add the terminal energy contribution:
This is algebraically the same weight as (8). When , every increment is zero; at , the full log weight is zero for any step size.
Algorithm 2 Gaussian-reference sampler with online log weights
Input: , energy , temperature , horizon , steps
, ,
,
for do
;
return
Weights are normalized in log space. At positive , the terminal energy term downweights high-energy endpoints; an energy evaluated as gives zero weight when the transition terms are finite. Resampling requires at least one finite log weight, and a batch with none must be treated as a failed refresh rather than normalized. This numerical issue is separate from the validity of the exact weight formula.
Justification of the score formula
The regression target. Fix and let , with . Assume that is differentiable and that differentiation under the integrals and the integration by parts below are valid. The Gaussian kernel satisfies . Consequently, for and ,
Since , integration by parts gives
The final equality follows because the conditional density of given is proportional to . This is the relative-score identity used in (4) [6, 32]. Least-squares regression learns the conditional mean of its target. Thus, when for , the population minimizer with regression target is . At , both are zero. Finite-buffer sampling and the target clipping used in the particle experiments are approximations to this population objective, not part of the identity.
The reverse sampler. The forward Ornstein–Uhlenbeck process has drift and noise coefficient . Its standard reverse-time formula [36], written with sampling time increasing from to , has drift
With the exact score and initial distribution , this process ends at . The explicit perturbation approaches the Gaussian reference as increases, motivating initialization from . At finite , this replacement is an approximation. The discrete importance weights in (8) remain valid for the numerical sampler regardless of this initial mismatch or score error.
Experimental details
Target definitions and reference distributions
Gaussian mixture model (GMM). We consider a synthetic target on with independent coordinates, each following a two-component Gaussian mixture:
where , , and . We use and , keeping the per-coordinate parameters fixed so that the two settings differ only in dimension. This construction yields well-separated modes with unequal probability masses. The main challenge is to explore these modes and recover their relative masses despite the low-density regions separating them. Exact independent samples can be obtained by sampling the mixture component and Gaussian noise independently for each coordinate.
4-particle double-well (DW-4). DW-4 is an interacting-particle benchmark for sampling multimodal Boltzmann distributions [26]. A configuration consists of four particles , represented by . The target density is
where and is the physical inverse temperature. We set , , , , and . Since the energy is translation-invariant, we impose and define the density with respect to Lebesgue measure on this six-dimensional subspace. The two preferred pairwise distances give rise to competing particle arrangements and multiple energy minima. The main sampling challenge is to cross the energy barriers separating these arrangements. As reference samples, we use the MCMC samples of Klein et al. [25], obtained from the DEM repository of Akhound-Sadegh et al. [1].
13-particle Lennard–Jones (LJ-13). LJ-13 models a cluster of thirteen particles , with configurations represented by . The Lennard–Jones potential models short-range repulsion and longer-range attraction between particles and provides a standard benchmark for sampling atomic cluster configurations. Köhler et al. [26]. We use the confined, regularized target
where . We set , , , , and the physical inverse temperature . The harmonic term confines the cluster, while regularizes the repulsive interaction near particle collisions. We impose and define the density with respect to Lebesgue measure on the resulting 36-dimensional subspace. The main sampling difficulties are transitions between metastable cluster configurations and the steep energy gradients associated with short-range repulsion. As for DW-4, the reference samples are the MCMC samples of Klein et al. [25], obtained from the DEM repository of Akhound-Sadegh et al. [1].
For the particle targets and the Gaussian reference, the relative energy used by the sampler is , up to an additive constant, with in these experiments.
Alanine Dipeptide. We sample the 22-atom ACE–ALA–NME molecule at 300 K in implicit solvent, using the AMBER ff14SB force field and OBC1 solvent model of Akhound-Sadegh et al. [2]. The target density is proportional to in the L-alanine sector. Its backbone angles distinguish conformations with different equilibrium populations. We compare with the 10,000 frames in the 300 K molecular-dynamics test split used by Akhound-Sadegh et al. [2].
SMC and HMC operate in 60 internal coordinates: 21 bonds, 20 angles and 19 torsions. The reference has 42 standard Gaussian coordinates for transformed bonds, angles and the chirality-defining improper torsion, and 18 independent uniform circular coordinates for the remaining torsions. The target density in these coordinates includes the coordinate Jacobian and the conformational volume factor. We wrap circular coordinates after each HMC position update. For diffusion training, we reconstruct Cartesian positions, subtract their center of mass, apply a random rotation and divide by the coordinate standard deviation of the final SMC population. The tempering reference in Cartesian space is the distribution induced by this map. The forward diffusion still uses Gaussian noise. Writing for the standard Gaussian density on the centered Cartesian subspace, the path weight in (8) uses the endpoint factor in scaled Cartesian coordinates. This reduces to when .
Neural network architecture
All experiments use a single network shared across the whole temperature ladder, conditioned on both the diffusion time and the temperature parameter . The architecture is chosen to match the structure of each target. We use a residual MLP for the Gaussian mixtures and an equivariant graph network for DW-4 and LJ-13, and a diffusion transformer for Alanine Dipeptide. On centered DW-4 and LJ-13 configurations, the target is unchanged by permutations of identical particles, rotations, and reflections.
Gaussian mixtures. The relative score network is a residual MLP with input , where denotes an -dimensional sinusoidal embedding with log-spaced frequencies in for and for . A linear layer maps the input to width 256, followed by three pre-norm residual blocks (RMSNorm, Dense, ELU, Dense) and a linear output layer. The output weights are initialized from with , and the output bias is set to zero, so that at initialization, consistent with the zero relative score at (§ 3.1). The network has parameters for and for .
Particle systems. For DW-4 and LJ-13, is an -equivariant graph neural network (EGNN) (Satorras et al. [34]) on the fully connected particle graph, with 5 layers of width 128, SiLU activations, attention-gated messages, and residual node updates. Each node receives as a scalar feature together with an 8-dimensional sinusoidal embedding of (frequencies in ), and each edge carries the squared inter-particle distance of the input configuration. Coordinate updates are bounded by tanh, with a per-layer range of 3. The output is the total coordinate displacement with its mean removed. It is therefore equivariant to permutations, rotations, and reflections, invariant to translations, and lies in the zero center-of-mass subspace (Appendix § 7.3.§ 3). The network has parameters.
Alanine Dipeptide. We use the Cartesian diffusion transformer of Akhound-Sadegh et al. [2]. A linear layer embeds each atom’s three coordinates into a token of width 192. Six transformer blocks with six attention heads are conditioned on the noise level and inverse temperature through a 64-dimensional embedding. The output layer is initialized to zero. We scale its prediction by , remove the center of mass and clip each component to . The network has 3,179,267 parameters. Training uses dropout 0.1 and fresh random rotations applied jointly to each configuration and its stored gradient. For the TSM term, we clip the clean relative score per atom at the 99.9th percentile of its norms in the initial buffer at that temperature.
Projection onto the zero center-of-mass subspace
For DW-4 and LJ-13, we represent a configuration of particles in as . We remove its center of mass by subtracting the mean particle position:
Equivalently, , where
is the orthogonal projection onto . This subspace has dimension . All diffusion sampling and density calculations below are understood on this space. The Gaussian reference is standard in an orthonormal coordinate system on the subspace; in ambient coordinates its covariance is .
For DW-4, the physical energy depends only on pairwise distances. Fixing the center of mass removes the unconstrained translation and makes the target a distribution over internal configurations. For LJ-13, the pair interaction is also translation-invariant, but the full potential is not, because it includes harmonic confinement. The confinement separates as
Thus the LJ-13 density factors into a center-of-mass term and an internal-coordinate term. Centering a draw preserves its internal-coordinate distribution; it does not assert translation invariance of the full confined target.
HMC steps. We use preconditioned HMC during initialization and training. Retained SMC particles are centered before entering the replay buffers; during training, the same centering is applied after each HMC move. The target-invariance claim requires the proposal and acceptance calculation to be consistent with the centered configuration space; projection alone does not establish it.
Diffusion steps. We center the Gaussian noise used in both training and sampling. For a centered buffer sample , the perturbation is
The score network also returns a centered vector, so sampling takes the form
Starting from a centered reference draw, all states remain centered. The learned and reference transitions have the same covariance in ambient coordinates, or in coordinates on the subspace. Their Gaussian normalizers therefore cancel, and the log-weight formula in Appendix § 7.1.§ 3 applies with the projected noise.
Reported metrics
Gaussian mixture models and Particle systems
For the Gaussian mixtures, DW-4, and LJ-13, we use the following discrepancies: Distance-, Energy-, and Geometric-, as considered in the evaluation of PITA [2]. Lower values indicate better agreement with the reference samples.
Distance-. Let denote a vector of summary statistics. For the Gaussian mixture targets, consists of the coordinates of ; for DW-4 and LJ-13, it consists of all pairwise interparticle distances. Given generated samples and reference samples , we pool these statistics into the one-dimensional empirical distributions
and report
The one-dimensional 2-Wasserstein distance admits the quantile representation [31]
where and are the corresponding quantile functions. Distance- measures agreement in the pooled coordinate or distance distribution, but does not capture dependencies among statistics within each configuration.
Energy-. Following the energy-distribution evaluation used by Akhound-Sadegh et al. [2], we compare the empirical distributions of the energy :
This metric measures whether generated samples reproduce the reference energy distribution, complementing Distance-. For both one-dimensional metrics, equal-sized scalar sample sets are compared exactly by sorting and computing the root mean squared difference. For unequal sizes, we approximate the quantile integral using 1000 uniformly spaced midpoint quantiles. The current evaluation excludes non-finite scalar values before either calculation. These metrics therefore do not penalize numerical failures: when exclusions occur, they describe only the retained finite subset and can understate the error of the full sample set.
**Geometric-\mathcal{W}_2$ compares pooled statistics and so cannot tell whether individual configurations are correct. For DW-4 and LJ-13, we therefore also compare their whole configurations. Two configurations that differ only by relabelling identical particles, or by a rotation or reflection, describe the same physical state, so we measure the distance between them after aligning out these symmetries. For configurations ( particles in dimensions, centred at the origin), the squared distance is
where ranges over particle permutations and over rotations and reflections. Geometric- is the 2-Wasserstein distance between the generated and reference samples with as the ground distance.
We approximate the joint alignment: following Köhler et al. [26], we match particles by a linear assignment and then fit a rotation or reflection to the matched coordinates by least squares (orthogonal Procrustes). The resulting squared distance is an upper bound on . We use these approximate alignment costs for the full sets of 10,000 generated and 10,000 reference samples. The outer transport problem is solved exactly for this cost matrix; the configuration alignment remains approximate.
Molecular system
For Alanine Dipeptide we use the 10,000-frame molecular-dynamics (MD) test split as the reference. SSS produces 10,000 diffusion endpoints and 4,096 buffer draws. We discard configurations with non-finite energy before evaluation. Both energy discrepancies use reduced energies and exact one-dimensional empirical transport, including when sample sizes differ. Energy- is defined above.
Energy-. For generated configurations and reference configurations , we compute
The ground distance is the absolute difference between two reduced energies. The metric measures the average displacement needed to match the two energy distributions.
T-. We represent each configuration by its backbone angles . For angle pairs and , we use the squared ground distance
We compute the 2-Wasserstein distance between the empirical distributions of generated and reference angle pairs using this ground distance and uniform sample weights. Measuring each difference around the circle keeps angles just across the periodic boundary close together.
Other metrics in the PITA benchmark. We omit Rama-KL because its histogram estimate depends strongly on sample size. Even between MD samples, our estimated floor is about 1.5 at 10,000 samples and 3.0 at 4,096, the two sizes used here. The published benchmark does not specify the bins, smoothing constant or sample count needed to match its estimate. We also omit TICA- and TICA-. Their projection is learned from pairs of MD frames 100 steps apart. The released 300 K splits are shuffled and have near-zero frame correlations at lags 1–100, so they do not preserve the information needed to reproduce that projection.
Numerical settings and baselines reimplementation
We list here the settings needed to reproduce our results. The numerical diffusion update uses the step size of each grid interval. The SMC particle count also sets the replay-buffer capacity , and counts all temperatures in the adaptive sweep, including (Appendix § 7.1.§ 1). The fixed training lengths below were selected using validation performance, as were the baseline training lengths. We report the cost of these final checkpoints, rather than time to a shared error tolerance or a formal convergence threshold.
GMM16 (). The diffusion horizon is . The SMC sweep uses particles, ESS threshold , and 3 HMC steps per temperature, giving for seeds 0, 1, and 2, respectively. Training runs for 10 outer iterations. Each iteration generates proposals per temperature with sampler steps, applies HMC steps after resampling, and then takes Adam steps. For final evaluation, we generate fresh samples with steps and compare them with exact draws from the target.
GMM50 (). The diffusion horizon is . The per-coordinate target is the same as in GMM16, but the higher dimension calls for more MCMC effort at initialization and longer training. The SMC sweep uses , , and 50 HMC steps per temperature, giving for all three seeds. Training runs for 10 outer iterations with , , , and . Evaluation follows GMM16, with fresh samples, , and exact reference draws.
DW-4 (). The diffusion horizon is . The SMC sweep uses , , and 10 HMC steps per temperature, giving for all three seeds. Training runs for 10 outer iterations with , , , and . To keep training stable under large energy gradients, we cap the norm of the regression target at 20 and clamp the network output to . For final evaluation, we generate fresh samples with steps.
LJ-13 (). The diffusion horizon is . The SMC sweep uses , , and 200 HMC steps per temperature, giving for all three seeds. Training runs for 50 outer iterations with , , , and . We cap the norm of the regression target at 150 per particle and clamp the network output to . For final evaluation, we generate fresh samples with steps.
Alanine Dipeptide. The diffusion horizon is . SMC uses 4,096 particles, an ESS threshold of 0.8, and 50 periodic HMC moves per level in internal coordinates. All three seeds give 40 levels including the reference. We recalibrate each step size on its final SMC population for use during training. Each of 100 outer iterations generates 1,024 paths per active level with 100 diffusion steps, applies 50 HMC moves, and takes 500 optimizer steps using 64 samples per buffer. We write a new block only if it has a finite path weight and the path-weight ESS is at least 1% of the batch size. Final evaluation uses 1,000 diffusion steps.
Alanine Dipeptide training loss. We draw log-uniformly on . For the prediction and the two regression targets in § 3.1, we minimize
We project the noise onto the zero center-of-mass subspace and clip the clean relative score in the TSM target as described in Appendix § 7.3.§ 2.
Temperature-based baselines. We reimplement PITA and TA-BG and report our runs on GMM16, GMM50, and DW-4. On LJ-13 and Alanine Dipeptide, both methods diverged on our hardware. For those two targets we use their published discrepancies from Akhound-Sadegh et al. [2]; local timing and evaluation budgets are unavailable for those entries. On the three successful targets, both methods use the positive inverse temperatures selected by the SSS warm-up for each seed. On both mixtures, each of PITA’s two networks uses the same residual MLP as SSS. On DW-4, each uses PITA’s EGNN with its width adjusted to within 0.8% of the SSS parameter count. PITA therefore has about twice as many parameters in total. TA-BG keeps its spline flow, with conditioner widths chosen so that the entire flow matches SSS within 0.9%.
Both baselines use the same preconditioned HMC kernel as SSS. We allocate the same nominal number of particle proposals as SSS uses in its warm-up and training rejuvenation combined, distributed across annealing transitions. PITA applies HMC after each transition. TA-BG applies it after importance resampling and before fitting the flow at the next temperature.
Adjoint Sampling. We reimplement AS (Havens et al. [20]) on all five targets and use an HMC-augmented version for the reported comparisons. For each seed, the final-temperature population from SSS’s SMC initialization seeds a uniformly weighted FIFO buffer. AS trains only at . Before each new batch enters the buffer, we apply the HMC kernel tuned by SMC at that temperature, with the same number of moves per particle as in SSS training. AS uses neither path-space importance resampling nor temperature-bridge weights.
We use SSS’s OU process in place of AS’s variance-exploding process, so the endpoint regression target coincides with the SSS target-score label at . For each target and seed, we match the network, time grid, optimizer, total diffusion paths, optimizer steps and total minibatch size. The AS buffer has the capacity of one SSS buffer. Including initialization, AS used 0.005–2% fewer energy-gradient evaluations than SSS.
Computational cost
Table Table 4 gives evaluation counts for the benchmark comparisons. Table Table 3 reports the available wall-clock measurements. Counts vary across seeds because the number of temperatures and the work needed to tune the kernel can vary. Thus, the reported runtimes and budget counts are averaged over 3 random seeds that we ran to produce the results. For PITA, TA-BG, and SSS, they use the same temperature ladder, so we run the methods until there is no improvement on the metrics or training loss. For AS, it does not contain any temperature setting, so we match the energy-evaluation of the runs with those in SSS. This will allow a much larger budget to be spent on AS compared to SSS at the target temperature .
| Target | Method | Energy (millions) | Gradient (millions) |
| GMM16 | SSS (ours) | 85.7 | 41.7 |
| PITA | 411.6 | 368.4 | |
| TA-BG | 86.9 | 44.9 | |
| AS | Matched SSS | Matched SSS | |
| GMM50 | SSS (ours) | 230.5 | 114.1 |
| PITA | 551.2 | 438.6 | |
| TA-BG | 227.0 | 115.3 | |
| AS | Matched SSS | Matched SSS | |
| DW-4 | SSS (ours) | 50.2 | 24.5 |
| PITA | 374.6 | 350.7 | |
| TA-BG | 98.0 | 73.8 | |
| AS | Matched SSS | Matched SSS | |
| LJ-13 | SSS (ours) | 9,077 | 3,038 |
| AS | Matched SSS | Matched SSS | |
| Alanine Dipeptide | SSS (ours) | 420.2 | 208.1 |
| AS | Matched SSS | Matched SSS |
Table 4. Energy and energy-gradient evaluations, averaged over three seeds. AS used matched SSS evaluation budgets on all five targets. Local budgets for TA-BG and PITA are unavailable on LJ-13 and Alanine Dipeptide.
Discussion on practical implementation
Initialization. One adaptive SMC pass selects the temperature ladder, seeds the buffers, and tunes a local kernel at each positive temperature. Training reuses these choices; for Alanine Dipeptide we recalibrate the step sizes on the final SMC populations. A stronger initialization costs more target evaluations up front, but can preserve modes that later local moves rarely reach. Initial mode weights need not be exact. In the selected eccentric-mixture run in Figure 4, SSS brought an underrepresented mode from 0.23 to 0.49; when the smaller initial population lost that mode, training did not recover it. This distinction helps explain why increasing the particle count matters at a bottleneck and has little effect on corrected samples once DW-4 is adequately covered.

Figure 4. Marginal densities of at for the selected seed, starting from SMC populations of 1,000 (left) and 80,000 (right) particles. Each panel compares the initial SMC population, corrected SSS samples, and exact samples. Insets report the corresponding far-mode shares, with exact target mass 0.50.
Importance-weight variance. The SMC ladder controls the change between adjacent targets through a bridge ESS criterion. Path weights depend on the learned proposal and can still concentrate even when adjacent targets overlap well. Early in training, the buffers retain SMC particles while fresh batches gradually replace them. At each refresh we resample the current paths, rejuvenate the endpoints, and normalize the new bridge weights within the batch, so path weights are never multiplied across outer iterations. Within a path, we use finer time steps near the endpoint, where the low-noise updates can contribute substantial weight variability. The grid is exponential on the mixtures, DW-4, and Alanine Dipeptide, and quadratic on LJ-13. These choices still leave few distinct endpoints when path ESS is low. On LJ-13 and Alanine Dipeptide, the final buffer gives lower energy discrepancies than resampling fresh paths alone.
Rejuvenation length. When resampling produces many copies of a few endpoints, longer chains give those copies more opportunity to explore their local basin. Each additional move costs target evaluations. In a separate LJ-13 chain-length experiment, increasing rejuvenation from 20 to 100 and then 2,000 moves reduced buffer Energy- from 2.31 to 0.37 and then 0.27. For fresh importance-corrected samples, the corresponding errors were 0.74, 0.40, and 0.36. Most of the observed reduction came in the first increase. Longer local chains can be worthwhile when duplicated particles are the main difficulty; when a whole mode is missing, allocating more particles to initialization can be more effective.
Additional baselines
In this Section, we consider two additional baselines for SSS: Initial sampler and Long run SMC.
Initial sampler
Before the first training update, the final buffer contains the population from the adaptive SMC sweep. We evaluate this population on the same three seeds and against the same reference samples as the trained SSS outputs in Table 1. Table 5 compares the four sample sets.
Table 5: Initial SMC population and SSS samples after training. Entries are mean standard deviation over three seeds, rounded to two decimal places. For each target and metric, bold marks the lowest displayed mean and <u>underlining</u> marks the second-lowest distinct mean. Equal displayed means share a rank. A dash denotes an inapplicable metric and denotes a mean above .
| Target | Samples | Distance- | Energy- | Geometric- |
| GMM16 () | SMC (initial) | 0.37 0.15 | 0.61 0.41 | — |
| SSS (raw) | 0.13 0.09 | 1.15 0.12 | — | |
| SSS (corrected) | 0.01 0.01 | 0.04 0.00 | — | |
| SSS (buffer) | <u>0.04</u> 0.01 | <u>0.05</u> 0.02 | — | |
| GMM50 () | SMC (initial) | 0.57 0.12 | 3.93 1.23 | — |
| SSS (raw) | <u>0.17</u> 0.03 | 4.30 0.81 | — | |
| SSS (corrected) | 0.05 0.02 | 0.32 0.06 | — | |
| SSS (buffer) | 0.21 0.04 | <u>1.03</u> 0.20 | — | |
| DW-4 () | SMC (initial) | 0.11 0.12 | 0.17 0.02 | 0.38 0.19 |
| SSS (raw) | 0.23 0.07 | 0.92 0.31 | 0.52 0.14 | |
| SSS (corrected) | <u>0.19</u> 0.01 | <u>0.11</u> 0.03 | <u>0.48</u> 0.01 | |
| SSS (buffer) | <u>0.19</u> 0.01 | <u>0.11</u> 0.01 | <u>0.48</u> 0.02 | |
| LJ-13 () | SMC (initial) | 0.00 0.00 | 0.23 0.06 | 1.46 0.00 |
| SSS (raw) | 0.04 0.03 | 1.50 0.04 | ||
| SSS (corrected) | 0.03 0.03 | 1.41 1.73 | 1.63 0.06 | |
| SSS (buffer) | <u>0.01</u> 0.00 | <u>0.32</u> 0.06 | <u>1.47</u> 0.00 |
Table 5.
On both Gaussian mixtures, corrected SSS samples have smaller mean distance and energy errors than the initialization. For example, on GMM50 the distance error falls from 0.57 to 0.05 and the energy error from 3.93 to 0.32. On DW-4, correction and buffering reduce the mean energy error, while the initial SMC population has lower mean distance and geometric errors. On LJ-13, initialization already gives the lowest mean discrepancies; the final buffer comes close, but training does not improve accuracy at . The trained model still provides a reusable way to generate fresh proposals at any temperature on the ladder.
Matched Energy budget SMC
We give a single SMC run the particle count, tempering rule, and preconditioned HMC kernel of the SSS warm-up. We increase the number of moves per temperature () to match the energy-evaluation budget of the entire SSS run, including initialization and training. This asks how much accuracy the same energy budget buys when it is spent entirely on SMC. On GMM16, the buffer reduces the mean distance error from 0.141 to 0.041 and the energy
error from 0.188 to 0.048. Both errors are roughly halved on GMM50 (Table 6). Once the mixture modes separate, extra local moves can explore each occupied mode while doing little to adjust how many particles represent it. Fresh proposals and importance correction give SSS another opportunity to change those proportions. The particle systems give a different result. On DW-4, SMC has lower distance and geometric errors, while the buffer has a slightly lower energy error. On LJ-13, SMC has lower means on all three metrics, with geometric errors close at 1.461 and 1.465.
Table 6: Final SMC samples and SSS buffer samples (mean standard deviation over three seeds). Bold marks the lower mean for each target and metric.
| Target | Samples | Distance- | Energy- | Geometric- |
| GMM16 | SMC () | 0.141 0.038 | 0.188 0.062 | — |
| SSS (buffer) | 0.041 0.009 | 0.048 0.016 | — | |
| GMM50 | SMC () | 0.383 0.128 | 2.240 1.020 | — |
| SSS (buffer) | 0.208 0.037 | 1.025 0.201 | — | |
| DW-4 | SMC () | 0.081 0.033 | 0.124 0.024 | 0.312 0.041 |
| SSS (buffer) | 0.194 0.010 | 0.105 0.014 | 0.479 0.016 | |
| LJ-13 | SMC () | 0.005 0.001 | 0.262 0.044 | 1.461 0.001 |
| SSS (buffer) | 0.006 0.002 | 0.324 0.063 | 1.465 0.003 |
Table 6.
Amortizing fixed SMC samples. On the Gaussian mixtures, we also train the same temperature-conditioned diffusion on fixed SMC samples at each temperature. We keep the target score-matching loss, optimizer, and training schedule of SSS, and turn off buffer refreshes. Storing the energy gradients adds at most about 1% to the SMC energy-evaluation count. We compare fresh raw samples from both models. On GMM50, refreshing the data during training reduces mean distance error from 0.398 to 0.166. On GMM16, the difference is smaller and the energy errors are similar.
| Target | Sampler | Distance- | Energy- |
| GMM16 | SMC + amortisation | 0.188 ± 0.042 | 1.166 ± 0.106 |
| SSS (raw) | 0.130 ± 0.091 | 1.153 ± 0.121 | |
| GMM50 | SMC + amortisation | 0.398 ± 0.125 | 5.462 ± 0.873 |
| SSS (raw) | 0.166 ± 0.034 | 4.304 ± 0.807 |
Table 7. Raw diffusion samples after training on fixed SMC populations or with SSS (mean ± standard deviation over three seeds). Bold marks the lower mean for each target and metric.
Additional ablation studies
Question 1 (Keeping a fixed buffer). Does fixing help by preserving an accurate starting population, and how much does an error in that buffer affect the final samples?
On DW-4, the first positive inverse temperature is close to the reference (). We compare keeping fixed with refreshing it from corrected model samples or from fresh reference draws reweighted to . We also fix it at deliberately inaccurate standard Gaussian draws. Even in the last case, where the first-buffer discrepancy rises to 0.667, the final corrected-sample errors remain close to the default (Table 8). Refreshing from the model gives slightly larger mean errors on all three metrics. Drawing fresh reference samples gives a similar distance error to the fixed buffer. These results support keeping the first buffer fixed on DW-4, where the first rung is already close to the reference.
| Distance- | Energy- | Geometric- | ||
| Fixed (ours) | 0.192 ± 0.010 | 0.109 ± 0.034 | 0.482 ± 0.014 | 0.008 |
| Self-refresh | 0.208 ± 0.015 | 0.126 ± 0.025 | 0.507 ± 0.024 | 0.007 |
| Reference bridge | 0.189 ± 0.049 | 0.138 ± 0.014 | 0.484 ± 0.082 | 0.008 |
| Bad anchor | 0.196 ± 0.023 | 0.122 ± 0.015 | 0.486 ± 0.037 | 0.667 |
Table 8. First-buffer ablation on DW-4. The first three metrics summarize corrected samples over three seeds. The final column measures the discrepancy between the first buffer and .
Question 2 (Choosing the number of SMC particles). How sensitive is SSS to the particle count when initialization covers the important configurations? What changes when a mode becomes rare along the tempering path?
On DW-4, SMC already covers all the modes at the smallest particle count tested. Corrected SSS errors change little as increases from 1,000 to 81,920. For instance, the corrected geometric error stays between 0.48 and 0.51 across the sweep (Table 9). This suggests that the benefit of extra particles depends on how well initialization covers the target. The SMC warm-up took 10–13 seconds at both endpoints on our GPU.
| Samples | Distance- | Energy- | Geometric- | |
| SSS (corrected) | 1,000 | 0.180 0.010 | 0.100 0.030 | 0.480 0.010 |
| 3,000 | 0.200 0.010 | 0.140 0.030 | 0.490 0.010 | |
| 5,000 | 0.200 0.010 | 0.120 0.010 | 0.490 0.010 | |
| 10,000 | 0.210 0.020 | 0.140 0.030 | 0.510 0.030 | |
| 20,000 | 0.200 0.020 | 0.120 0.020 | 0.500 0.030 | |
| 50,000 | 0.190 0.010 | 0.130 0.030 | 0.490 0.020 | |
| 81,920 | 0.190 0.010 | 0.110 0.030 | 0.480 0.010 | |
| SMC | 1,000 | 0.350 0.070 | 0.450 0.250 | 1.140 0.410 |
| 3,000 | 0.330 0.180 | 0.140 0.010 | 0.910 0.330 | |
| 5,000 | 0.270 0.170 | 0.130 0.030 | 0.690 0.280 | |
| 10,000 | 0.350 0.160 | 0.200 0.120 | 0.850 0.310 | |
| 20,000 | 0.400 0.200 | 0.090 0.050 | 0.880 0.370 | |
| 50,000 | 0.070 0.030 | 0.100 0.040 | 0.330 0.070 | |
| 81,920 | 0.110 0.120 | 0.180 0.020 | 0.380 0.190 |
Table 9. Effect of SMC particle count on DW-4. Entries are mean standard deviation.
To study the second question, we use the Eccentric GMM. In dimension , its target density is
where is the vector of ones. We call the component centered at the far mode because its center lies farther from the center of the reference distribution . Quadrature gives a far-mode mass of only near , although its mass at is 0.5. We ran SMC over 100 seeds at each particle count, matching the energy budget of 80,000 particles with ten HMC moves per temperature. Smaller populations received more HMC moves. Table 10 shows that increasing the particle count from 1,000 to 80,000 raised the fraction of runs retaining the far mode from 3% to 85%. At this bottleneck, longer local chains did little to compensate for a small population.
| Particles | HMC moves per temperature | Runs retaining the far mode |
| 1,000 | 1,000 | 3% |
| 3,000 | 330 | 13% |
| 5,000 | 195 | 25% |
| 10,000 | 97 | 36% |
| 20,000 | 47 | 50% |
| 50,000 | 17 | 71% |
| 80,000 | 10 | 85% |
Table 10. Far-mode retention in the eccentric Gaussian mixture at a fixed SMC energy budget. Each row summarizes 100 runs.
We then followed SSS training for one seed chosen to illustrate the two outcomes (Figure 4). With 80,000 particles, SMC retained the far mode but assigned it mass 0.23. Within a few outer iterations SSS brought that share to 0.49, close to the exact 0.50. With 1,000 particles, SMC lost the mode and SSS cannot recover it.
Question 3 (Choosing the MCMC kernel). How much does the choice of MCMC kernel affect SSS?
We replace the kernel in both the SMC warm-up and the training rejuvenation steps. Each kernel uses the same empirical covariance preconditioner and number of moves per temperature. We tune the step size by bisection, targeting acceptance rates of 0.8–0.95 for one-step HMC with partial momentum refresh, 0.4–0.65 for MALA, and 0.2–0.35 for random-walk Metropolis–Hastings (RWMH).
Kernel choice has relatively little effect on GMM16 and DW-4. On GMM50, the two gradient-based kernels give much smaller energy errors than RWMH. On LJ-13, HMC gives the lowest corrected-sample errors on all three metrics, while buffer geometric errors are similar across kernels (Table 11 and Table 12). The gap is larger for fresh corrected proposals than for the stored LJ-13 population. This distinction matters when the goal is to train a sampler for repeated use.
Table 11: Kernel comparison for SSS corrected samples (mean standard deviation over three seeds). For each target and metric, bold marks the lowest displayed mean and underlining marks the second-lowest distinct mean.
| Target | Kernel | Distance- | Energy- | Geometric- |
| GMM16 | HMC (default) | 0.005 0.005 | 0.040 0.004 | — |
| MALA | 0.017 0.005 | 0.031 0.005 | — | |
| RWMH | <u>0.016</u> 0.014 | <u>0.033</u> 0.007 | — | |
| DW-4 | HMC (default) | <u>0.192</u> 0.010 | 0.109 0.034 | <u>0.482</u> 0.014 |
| MALA | 0.191 0.004 | 0.137 0.006 | 0.477 0.008 | |
| RWMH | 0.200 0.012 | <u>0.118</u> 0.007 | 0.495 0.023 | |
| GMM50 | HMC (default) | <u>0.048</u> 0.022 | 0.321 0.059 | — |
| MALA | 0.033 0.023 | <u>0.369</u> 0.199 | — | |
| RWMH | 0.360 0.071 | 7.731 5.996 | — | |
| LJ-13 | HMC (default) | 0.026 0.027 | 1.405 1.733 | 1.626 0.059 |
| MALA | 0.200 0.283 | 14.460 18.850 | 2.205 0.659 | |
| RWMH | <u>0.125</u> 0.189 | <u>8.320</u> 12.770 | <u>2.102</u> 0.758 |
Table 11.
Table 12: Kernel comparison for SSS buffer samples (mean standard deviation over three seeds). For each target and metric, bold marks the lowest displayed mean and underlining marks the second-lowest distinct mean. Equal displayed means share a rank.
| Target | Kernel | Distance- | Energy- | Geometric- |
| GMM16 | HMC (default) | 0.041 0.009 | 0.048 0.016 | — |
| MALA | 0.019 0.013 | 0.042 0.005 | — | |
| RWMH | <u>0.026</u> 0.028 | <u>0.044</u> 0.014 | — | |
| DW-4 | HMC (default) | 0.194 0.010 | 0.105 0.014 | 0.479 0.016 |
| MALA | <u>0.200</u> 0.006 | <u>0.114</u> 0.024 | <u>0.489</u> 0.007 | |
| RWMH | 0.204 0.013 | 0.135 0.007 | 0.490 0.020 | |
| GMM50 | HMC (default) | <u>0.208</u> 0.037 | <u>1.025</u> 0.201 | — |
| MALA | 0.136 0.124 | 0.589 0.650 | — | |
| RWMH | 0.430 0.322 | 6.069 1.520 | — | |
| LJ-13 | HMC (default) | 0.006 0.002 | 0.324 0.063 | <u>1.465</u> 0.003 |
| MALA | <u>0.012</u> 0.010 | 0.501 0.319 | 1.472 0.010 | |
| RWMH | <u>0.012</u> 0.001 | <u>0.409</u> 0.160 | 1.463 0.005 |
Table 12.
References
- [1]Tara Akhound-Sadegh, Jarrid Rector-Brooks, Avishek 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 Forty-first International Conference on Machine Learning, ICML 2024, Vienna, Austria, July 21-27, 2024, volume 235 of Proceedings of Machine Learning Research, pages 760–786. PMLR / OpenReview.net, 2024.
- [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 Advances in Neural Information Processing Systems 38: Annual Conference on Neural Information Processing Systems 2025, NeurIPS 2025, San Diego, CA, USA, December 2-7, 2025 / Mexico City, Mexico, November 30 - December 5, 2025, 2025.
- [3]Michael Samuel Albergo and Eric Vanden-Eijnden. NETS: A non-equilibrium transport sampler. In Forty-second International Conference on Machine Learning, ICML 2025, Vancouver, BC, Canada, July 13-19, 2025, volume 267 of Proceedings of Machine Learning Research. PMLR / OpenReview.net, 2025.arxiv.org/abs/2410.02711
- [4]Michael Arbel, Alexander G. de G. Matthews, and Arnaud Doucet. Annealed flow transport Monte Carlo. In Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, volume 139 of Proceedings of Machine Learning Research, pages 318–330. PMLR, 2021.arxiv.org/abs/2102.07501
- [5]Alexandros Beskos, Ajay Jasra, Nikolas Kantas, and Alexandre Thiery. On the convergence of adaptive sequential Monte Carlo methods. Ann. Appl. Probab., 26(2):1111–1146, 2016. ISSN 1050-5164. doi: 10.1214/15-AAP1113.DOI
- [6]Valentin De Bortoli, Michael J. Hutchinson, Peter Wirnsberger, and Arnaud Doucet. Target score matching. CoRR, abs/2402.08667, 2024. doi: 10.48550/ARXIV.2402.08667. URL https://doi.org/10.48550/arXiv.2402.08667.
- [7]Junhua Chen, Lorenz Richter, Julius Berner, Denis Blessing, Gerhard Neumann, and Anima Anandkumar. Sequential controlled Langevin diffusions. In The Thirteenth International Conference on Learning Representations, ICLR 2025, Singapore, April 24-28, 2025. OpenReview.net, 2025.arxiv.org/abs/2412.07081
- [8]Sanghyeok Choi, Sarthak Mittal, Víctor Elvira, Jinkyou Park, and Nikolay Malkin. Reinforced sequential Monte Carlo for amortised sampling. CoRR, abs/2510.11711, 2025. doi: 10.48550/ARXIV.2510.11711. URL https://doi.org/10.48550/arXiv.2510.11711.
- [9]Nicolas Chopin. A sequential particle filter method for static models. Biometrika, 89(3):539–551, 2002. ISSN 0006-3444. doi: 10.1093/biomet/89.3.539.DOI
- [10]Gavin E Crooks. Entropy production fluctuation theorem and the nonequilibrium work relation for free energy differences. Physical Review E, 60(3):2721, 1999.
- [11]Chenguang Dai, Jeremy Heng, Pierre E. Jacob, and Nick Whiteley. An invitation to sequential Monte Carlo samplers. J. Am. Stat. Assoc., 117(539):1587–1600, 2022. ISSN 0162-1459. doi: 10.1080/01621459.2022.2087659.DOI
- [12]Alexander G. de G. Matthews, Michael Arbel, Danilo Jimenez Rezende, and Arnaud Doucet. Continual repeated annealed flow transport Monte Carlo. In International Conference on Machine Learning, ICML 2022, 17-23 July 2022, Baltimore, Maryland, USA, volume 162 of Proceedings of Machine Learning Research, pages 15196–15219. PMLR, 2022.
- [13]Pierre Del Moral, Arnaud Doucet, and Ajay Jasra. Sequential Monte Carlo samplers. J. R. Stat. Soc., Ser. B, Stat. Methodol., 68(3):411–436, 2006. ISSN 1369-7412. doi: 10.1111/j.1467-9868.2006.00553.x.DOI
- [14]Manuel Dibak, Leon Klein, Andreas Krämer, and Frank Noé. Temperature steerable flows and Boltzmann generators. Physical Review Research, 4(4):L042005, 2022.
- [15]Arnaud Doucet, Will Grathwohl, Alexander G. de G. Matthews, and Heiko Strathmann. Score-based diffusion meets annealed importance sampling. In Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems 2022, NeurIPS 2022, New Orleans, LA, USA, November 28 - December 9, 2022, 2022.
- [16]David J Earl and Michael W Deem. Parallel tempering: Theory, applications, and new perspectives. Physical Chemistry Chemical Physics, 7(23):3910–3916, 2005.
- [17]Bradley Efron. Tweedie’s formula and selection bias. J. Am. Stat. Assoc., 106(496):1602–1614, 2011. ISSN 0162-1459. doi: 10.1198/jasa.2011.tm11181.DOI
- [18]Charles J. Geyer. Markov chain Monte Carlo maximum likelihood. In Computing Science and Statistics: Proceedings of the 23rd Symposium on the Interface, pages 156–163, 1991.
- [19]Walter R. Gilks and Carlo Berzuini. Following a moving target – Monte Carlo inference for dynamic Bayesian models. J. R. Stat. Soc., Ser. B, Stat. Methodol., 63(1):127–146, 2001. ISSN 1369-7412. doi: 10.1111/1467-9868.00280.
- [20]Aaron Havens, Benjamin Kurt Miller, Bing Yan, Carles Domingo-Enrich, Anuroop Sriram, Brandon Wood, Daniel Levine, Bin Hu, Brandon Amos, Brian Karrer, Xiang Fu, Guan-Horng Liu, and Ricky T. Q. Chen. Adjoint sampling: Highly scalable diffusion samplers via adjoint matching. arXiv preprint arXiv:2504.11713, 2025. doi: 10.48550/arXiv.2504.11713.
- [21]Koji Hukushima and Koji Nemoto. Exchange Monte Carlo method and application to spin glass simulations. Journal of the Physical Society of Japan, 65(6):1604–1608, 1996.
- [22]Michele Invernizzi, Andreas Kramer, Cecilia Clementi, and Frank Noé. Skipping the replica exchange ladder with normalizing flows. The Journal of Physical Chemistry Letters, 13(50):11643–11649, 2022.
- [23]Christopher Jarzynski. Nonequilibrium equality for free energy differences. Physical review letters, 78(14):2690, 1997.DOI
- [24]Ajay Jasra, David A. Stephens, Arnaud Doucet, and Theodoros Tsagaris. Inference for Lévy-driven stochastic volatility models via adaptive sequential Monte Carlo. Scand. J. Stat., 38(1):1–22, 2011. ISSN 0303-6898. doi: 10.1111/j.1467-9469.2010.00723.x.
- [25]Leon Klein, Andreas Krämer, and Frank Noé. Equivariant flow matching. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, 2023.
- [26]Jonas Köhler, Leon Klein, and Frank Noé. Equivariant flows: Exact likelihood generative learning for symmetric densities. In Proceedings of the 37th International Conference on Machine Learning, ICML 2020, 13-18 July 2020, Virtual Event, volume 119 of Proceedings of Machine Learning Research, pages 5361–5370. PMLR, 2020.arxiv.org/abs/2006.02425
- [27]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, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net, 2023.
- [28]Radford M Neal. Annealed importance sampling. Statistics and computing, 11(2):125–139, 2001.
- [29]Jerome P Nilmeier, Gavin E Crooks, David DL Minh, and John D Chodera. Nonequilibrium candidate Monte Carlo is an efficient tool for equilibrium simulation. Proceedings of the National Academy of Sciences, 108(45):E1009–E1018, 2011.DOI
- [30]Frank Noé, Simon Olsson, Jonas Köhler, and Hao Wu. Boltzmann generators: Sampling equilibrium states of many-body systems with deep learning. Science, 365(6457):eaaw1147, 2019.
- [31]Gabriel Peyré and Marco Cuturi. Computational optimal transport. With applications to data sciences. Found. Trends Mach. Learn., 11(5-6):1–262, 2018. ISSN 1935-8237. doi: 10.1561/2200000073.
- [32]Angus Phillips, Hai-Dang Dau, Michael John Hutchinson, Valentin De Bortoli, George Deligiannidis, and Arnaud Doucet. Particle denoising diffusion sampler. In Forty-first International Conference on Machine Learning, ICML 2024, Vienna, Austria, July 21-27, 2024, volume 235 of Proceedings of Machine Learning Research, pages 40688–40724. PMLR / OpenReview.net, 2024.arxiv.org/abs/2402.06320
- [33]Severi Rissanen, Ruikang Ouyang, Jiajun He, Wenlin Chen, Markus Heinonen, Arno Solin, and José Miguel Hernández-Lobato. Progressive tempering sampler with diffusion. In Forty-second International Conference on Machine Learning, ICML 2025, Vancouver, BC, Canada, July 13-19, 2025, volume 267 of Proceedings of Machine Learning Research. PMLR / OpenReview.net, 2025.arxiv.org/abs/2506.05231
- [34]Victor Garcia Satorras, Emiel Hoogeboom, and Max Welling. E(n) equivariant graph neural networks. In Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, volume 139 of Proceedings of Machine Learning Research, pages 9323–9332. PMLR, 2021.
- [35]Henrik Schopmans and Pascal Friederich. Temperature-annealed Boltzmann generators. In Forty-second International Conference on Machine Learning, ICML 2025, Vancouver, BC, Canada, July 13-19, 2025, volume 267 of Proceedings of Machine Learning Research. PMLR / OpenReview.net, 2025. URL https://proceedings.mlr.press/v267/schopmans25a.html.
- [36]Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021.
- [37]Robert H Swendsen and Jian-Sheng Wang. Replica Monte Carlo simulation of spin-glasses. Physical review letters, 57(21):2607, 1986.pubmed.ncbi.nlm.nih.gov/10033814
- [38]Charlie B. Tan, Joey Bose, Chen Lin, Leon Klein, Michael M. Bronstein, and Alexander Tong. Scalable equilibrium sampling with sequential Boltzmann generators. In Forty-second International Conference on Machine Learning, ICML 2025, Vancouver, BC, Canada, July 13-19, 2025, volume 267 of Proceedings of Machine Learning Research. PMLR / OpenReview.net, 2025.arxiv.org/abs/2502.18462
- [39]Francisco Vargas, Will Sussman Grathwohl, and Arnaud Doucet. Denoising diffusion samplers. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net, 2023.arxiv.org/abs/2302.13834
- [40]Francisco Vargas, Shreyas Padhy, Denis Blessing, and Nikolas Nüsken. Transport meets variational inference: Controlled Monte Carlo diffusions. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net, 2024.arxiv.org/abs/2307.01050
- [41]Pascal Vincent. A connection between score matching and denoising autoencoders. Neural computation, 23(7):1661–1674, 2011.DOI
- [42]Chenguang Wang, Xiaoyu Zhang, Kaiyuan Cui, Weichen Zhao, Yongtao Guan, and Tianshu Yu. Importance weighted score matching for diffusion samplers with enhanced mode coverage. CoRR, abs/2505.19431, 2025. doi: 10.48550/ARXIV.2505.19431.DOI
- [43]Yihang Wang, Lukas Herron, and Pratyush Tiwary. From data to noise to data for mixing physics across temperatures with generative artificial intelligence. Proceedings of the National Academy of Sciences, 119(32): e2203656119, 2022.DOI
- [44]Dawn B. Woodard, Scott C. Schmidler, and Mark Huber. Conditions for rapid mixing of parallel and simulated tempering on multimodal distributions. Ann. Appl. Probab., 19(2):617–640, 2009. ISSN 1050-5164. doi: 10.1214/08-AAP555.DOI
- [45]Hao Wu, Jonas Köhler, and Frank Noé. Stochastic normalizing flows. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual, 2020.arxiv.org/abs/2002.06707
- [46]Qinsheng Zhang and Yongxin Chen. Path integral sampler: A stochastic control approach for sampling. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net, 2022.arxiv.org/abs/2111.15141