- Rating
- 2(+1)
- Confidence
- 4
- Informativeness
- 1
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
The paper proposes corrected self-training across tempered distributions and instantiates it as Sequential Score Sampler (SSS). A single temperature-conditioned diffusion model is trained using weighted replay buffers at all temperatures. To refresh a buffer, the method generates proposals at the preceding temperature, corrects them using path-space importance weights, resamples and applies target-invariant MCMC, and then reweights the resulting population toward the next temperature. Experiments compare raw model samples, importance-corrected samples, and buffer samples against TA-BG and PITA.
Critical strengths
-
The method combines tractable importance correction with supervised score regression, avoiding evaluation of the diffusion model’s marginal likelihood.
-
Sharing one model across temperatures and caching energy gradients in replay buffers are useful practical features.
-
The paper includes ablations examining the roles of importance correction and MCMC rejuvenation.
Critical weaknesses
-
The incremental methodological contribution is insufficiently established. Progressive training across temperatures is closely related to PITA [2], PTSD [3], and TA-BG [1], while path-space correction is an established technique. Although the paper cites foundational work, the contribution of its particular combination remains unclear. Relevant additional context includes VT-DIS [4], which corrects diffusion proposals using trajectory-wise importance weights, and MAD-Path [5], which performs an augmented-path-space Metropolis correction. These methods differ from SSS, but illustrate why avoiding marginal likelihood evaluation through path-space correction is not itself a new principle. The paper should identify more precisely what new capability or insight its combination provides.
-
The advantage of separating proposal correction from temperature reweighting is not established. For a trajectory generated at , a direct importance weight targeting is
=w_k(y_{0:L})\exp\left[-(\beta_{k+1}-\beta_k)E(y_L)\right].$$ Correction and annealing can therefore be combined before a single resampling operation, followed by MCMC targeting $\pi_{\beta_{k+1}}$. SSS instead places MCMC at $\beta_k$ between the two weighting stages, which could affect exploration and finite-sample behavior. However, the paper does not establish when this ordering is preferable. The existing ablations remove correction or MCMC altogether rather than test this direct alternative. A controlled comparison under matched computational budgets would help justify the design. -
The experimental scope is limited. The benchmarks comprise factorized GMMs in 16 and 50 dimensions, DW-4, and LJ-13. Although the GMM has exponentially many modes, its independent-coordinate structure provides limited evidence about scaling to strongly coupled distributions. Evaluation on a more demanding system, such as LJ-55 or a molecular Boltzmann distribution, would provide substantially stronger evidence of practical applicability. The molecular experiments in Adjoint Sampling [6] illustrate relevant settings. The current results leave performance beyond relatively small benchmark systems unresolved.
-
SITA is a particularly relevant missing discussion. SITA [7] also uses progressive annealing and importance-weighted bootstrapping, employing an energy-based surrogate likelihood estimator to avoid expensive density calculations. This is directly relevant to the paper's discussion of PITA [2] and auxiliary energy models. Discussing the trade-off between surrogate likelihoods and exact discrete path-space weights would sharpen the positioning of SSS.
-
The comparison with modern neural samplers is narrow. The empirical comparison includes only TA-BG [1] and PITA [2], while the related-work discussion omits Adjoint Sampling [6], Adjoint Schrödinger Bridge Sampler [8], and Bridge Matching Sampler [9]. Their relationship to SSS should be discussed in terms of training objectives, sample reuse, correction mechanisms, and computational requirements. A comparison with an appropriate representative would help establish the method's practical contribution beyond the two selected annealing baselines.
-
The efficiency evaluation omits energy-function evaluation costs. Table 2 reports wall-clock time, including initialization, training, and sampling, but does not compare total energy and energy-gradient evaluations. These are important efficiency metrics for neural samplers, especially when the target energy or forces are expensive to evaluate. Wall-clock comparisons depend on hardware, implementation, and the relative costs of neural-network and energy evaluations, so they do not establish efficiency in this regime. This omission is particularly consequential because SSS uses SMC initialization and repeated HMC rejuvenation; Appendix C.5 specifies 2,000 HMC moves per particle per refresh on LJ-13. The comparison should account for energy and gradient evaluations across initialization, buffer updates, and final sample generation, including the benefit of cached gradients. Sample-quality curves against these evaluation budgets would make the efficiency claims more convincing.
References
[1] Henrik Schopmans and Pascal Friederich. Temperature-Annealed Boltzmann Generators. ICML, 2025. https://proceedings.mlr.press/v267/schopmans25a.html
[2] Tara Akhound-Sadegh et al. Progressive Inference-Time Annealing of Diffusion Models for Sampling from Boltzmann Densities. NeurIPS, 2025. https://arxiv.org/abs/2506.16471
[3] Severi Rissanen et al. Progressive Tempering Sampler with Diffusion. ICML, 2025. https://proceedings.mlr.press/v267/rissanen25a.html
[4] Fengzhe Zhang, Laurence I. Midgley, and José Miguel Hernández-Lobato. Efficient and Unbiased Sampling from Boltzmann Distributions via Variance-Tuned Diffusion Models. arXiv:2505.21005, 2025. https://arxiv.org/abs/2505.21005
[5] Han Chen, Sifan Liu, and Jun Yang. Markov Chain Monte Carlo with Diffusion Paths. arXiv:2607.11631, 2026. https://arxiv.org/abs/2607.11631
[6] Aaron J. Havens et al. Adjoint Sampling: Highly Scalable Diffusion Samplers via Adjoint Matching. ICML, 2025. https://proceedings.mlr.press/v267/havens25a.html
[7] Daniel Peñaherrera, Rishal Aggarwal, and David Ryan Koes. Scalable Inference-Time Annealing with Surrogate Likelihood Estimators. arXiv:2605.31498, 2026. https://arxiv.org/abs/2605.31498
[8] Guan-Horng Liu et al. Adjoint Schrödinger Bridge Sampler. NeurIPS, 2025. https://arxiv.org/abs/2506.22565
[9] Denis Blessing et al. Bridge Matching Sampler: Scalable Sampling via Generalized Fixed-Point Diffusion Matching. arXiv:2603.00530, 2026, version 1. https://arxiv.org/abs/2603.00530v1
Questions for authors
-
What benefit does correcting and rejuvenating at provide over directly weighting to and rejuvenating there? Can this be demonstrated under matched computational budgets?
-
Which aspect of the proposed combination provides an advantage beyond existing temperature-annealing and path-space correction techniques?
-
What are the total energy and energy-gradient evaluation costs for SSS and the baselines, broken down by initialization, training, and final sampling? Does SSS retain its advantage at matched energy-evaluation budgets?
6 comments
Answer from Authors to the Question 1 from Reviewer
Question 1. Ordering comparison
In the Direct ablation, trajectories generated at are resampled once using the combined weight
Endpoints are then rejuvenated with the HMC kernel adapted by SMC at and stored with uniform weights. Both variants use the same configuration and the same number of energy and gradient evaluations per outer step. Each cell reports the mean over three seeds for SSS / Direct, with the better mean in bold. Lower is better; denotes a mean error above .
| Target | Samples | Distance- | Energy- | Geometric- |
|---|---|---|---|---|
| GMM16 | Raw | 0.130 / 0.039 | 1.153 / 0.930 | NA |
| Corrected | 0.005 / 0.018 | 0.040 / 0.039 | NA | |
| Buffer | 0.041 / 0.011 | 0.048 / 0.037 | NA | |
| GMM50 | Raw | 0.166 / 0.132 | 4.304 / 4.850 | NA |
| Corrected | 0.048 / 0.129 | 0.321 / 1.340 | NA | |
| Buffer | 0.208 / 0.124 | 1.025 / 0.760 | NA | |
| DW4 | Raw | 0.227 / 0.089 | 0.924 / 1.590 | 0.523 / 0.318 |
| Corrected | 0.192 / 0.198 | 0.109 / 0.113 | 0.482 / 0.496 | |
| Buffer | 0.194 / 0.189 | 0.105 / 0.099 | 0.479 / 0.467 | |
| LJ13 | Raw | 0.041 / 0.104 | / | 1.499 / 1.501 |
| Corrected | 0.026 / 0.087 | 1.405 / 4.020 | 1.626 / 1.970 | |
| Buffer | 0.006 / 0.005 | 0.324 / 0.250 | 1.465 / 1.459 |
Normalized path-space ESS of the trained sampler at (SSS / Direct, higher is better).
| GMM16 | GMM50 | DW4 | LJ13 |
|---|---|---|---|
| 0.423 / 0.386 | 0.0077 / 0.0018 | 0.46 / 0.50 | 0.038 / 0.014 |
Corrected SSS samples have lower mean discrepancies on 9 of 10 target–metric comparisons. Direct has lower final-buffer means on all 10 comparisons, including a reduction from 0.041 to 0.011 in GMM16 distance error. SSS has higher path ESS on GMM16, GMM50, and LJ13; Direct has higher path ESS on DW4 (0.50 versus 0.46).
SSS applies the bridge to an approximately -distributed population after correction and rejuvenation. This matches the distribution used to calibrate SMC's bridge-ESS criterion. The bridge weights are retained for sampling training minibatches. Direct combines proposal correction and bridging before resampling, and applies MCMC at the destination temperature, including the final target.
Answer from Authors to the Question 2 from Reviewer
Question 2. Benefit of corrected self-training
We compare SSS with the same diffusion fitted to fixed SMC populations, retaining the architecture, target score-matching loss and training schedule. On GMM50, the fixed-data model's raw Distance- is , compared with for SSS. The improvement is present in fresh model draws, before final importance correction. The comparison is reported under Additional baselines in the revised appendix.
Response to Reviewer WEY2
We would like to thank you for your comments. While we respectfully disagree that all of your concerns are truly "critical weaknesses", we have considered them seriously and revised our manuscript accordingly. We clarified the contribution, added Alanine Dipeptide and an Adjoint Sampling baseline, and report energy and gradient evaluation counts. We tested the proposed alternative ordering under matched budgets.
Response to the main points
Point 1: Methodological contribution
The contribution is a general corrected self-training procedure for one temperature-conditioned proposal. At each intermediate temperature, proposals are importance-corrected and rejuvenated, then reweighted to refresh training data at the next temperature. One shared model learns jointly from these repeatedly refreshed populations. The procedure accommodates different proposal families, suitable fitting losses, initialization samplers and invariant MCMC kernels, provided valid importance weights are available. SSS instantiates the procedure with a diffusion trained by score regression and directly computable path weights, without an auxiliary likelihood model.
Path-space correction is established, as illustrated by VT-DIS and MAD-Path. Our contribution concerns its use within repeated training-data correction and joint fitting across temperatures. The revised manuscript states this distinction and separately evaluates raw proposals, corrected proposals and the final buffer.
Point 2: Correction and temperature reweighting
Both orderings give valid Monte Carlo constructions. SSS corrects and rejuvenates proposals at the temperature where they were generated, then applies the bridge to the adjacent target. The matched-budget comparison favors SSS for fresh importance-corrected proposals and Direct for final-buffer samples, as detailed in Question 1.
Point 3: Experimental scope
We added Alanine Dipeptide as a molecular Boltzmann target. Among the compared methods, raw SSS has the lowest mean angular discrepancy, while its buffer has the lowest energy discrepancies. Table 2 of the general response reports the full results and identifies the published PITA and TA-BG entries.
Point 4: Relation to SITA
SITA (Peñaherrera et al., 2026) progressively fine-tunes a flow using a learned surrogate likelihood for importance weights. Our procedure jointly trains a temperature-conditioned proposal from repeatedly refreshed populations across the ladder. SSS computes exact path-density ratios for its discrete sampler, without a surrogate likelihood. The methods share annealing and importance-weighted training, with different population updates and correction weights.
Point 5: Neural-sampler comparisons
We added an HMC-augmented Adjoint Sampling baseline using SSS's network, SMC initialization and matched computational settings. The general response reports its results; the revised appendix specifies the implementation.
Adjoint Schrödinger Bridge Sampler extends adjoint matching to arbitrary source distributions. Bridge Matching Sampler learns a stochastic transport through generalized fixed-point diffusion matching. These provide alternative fitting objectives for stochastic generators. SSS trains its generator on explicitly corrected populations that are repeatedly refreshed across temperatures.
Point 6: Energy and gradient evaluation costs
The table reports total energy and energy-gradient evaluations, in millions, averaged over 3 seeds.
Count variation reflects the adaptive temperature ladder and kernel tuning. On LJ13 and Alanine Dipeptide, we use published PITA and TA-BG results; local evaluation counts are unavailable for those entries.