acceptodds
Under review as a conference paper at ICLR 2027

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 is competitive in terms of mean discrepancy with the two baselines we considered, while taking less wall-clock time.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
diffusion samplers, adaptive tempering, importance sampling, score matching, Boltzmann generators

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_WEY2OpenReview ↗
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
  1. The method combines tractable importance correction with supervised score regression, avoiding evaluation of the diffusion model’s marginal likelihood.

  2. Sharing one model across temperatures and caching energy gradients in replay buffers are useful practical features.

  3. The paper includes ablations examining the roles of importance correction and MCMC rejuvenation.

Critical weaknesses
  1. 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.

  2. The advantage of separating proposal correction from temperature reweighting is not established. For a trajectory generated at βk\beta_k, a direct importance weight targeting βk+1\beta_{k+1} 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.
  3. 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.

  4. 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.

  5. 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.

  6. 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
  1. What benefit does correcting and rejuvenating at βk\beta_k provide over directly weighting to βk+1\beta_{k+1} and rejuvenating there? Can this be demonstrated under matched computational budgets?

  2. Which aspect of the proposed combination provides an advantage beyond existing temperature-annealing and path-space correction techniques?

  3. 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
AuthorsOpenReview ↗

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.

TargetMethodEnergy evals (M)Gradient evals (M)
GMM16SSS (ours)85.741.7
PITA411.6368.4
TA-BG86.944.9
ASMatched SSSMatched SSS
GMM50SSS (ours)230.5114.1
PITA551.2438.6
TA-BG227.0115.3
ASMatched SSSMatched SSS
DW4SSS (ours)50.224.5
PITA374.6350.7
TA-BG98.073.8
ASMatched SSSMatched SSS
LJ13SSS (ours)9,0773,038
ASMatched SSSMatched SSS
ALDPSSS (ours)420.2208.1
ASMatched SSSMatched SSS

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.

AuthorsOpenReview ↗

Answer from Authors to the Question 1 from Reviewer

Question 1. Ordering comparison

In the Direct ablation, trajectories generated at βk\beta_k are resampled once using the combined weight

wk(y0:L)⋅exp⁡[−(βk+1−βk) E(yL)].w_k(y_{0:L}) \cdot \exp\left[-(\beta_{k+1} - \beta_k)\,E(y_L)\right].

Endpoints are then rejuvenated with the HMC kernel adapted by SMC at βk+1\beta_{k+1} 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; ∗\ast denotes a mean error above 10410^4.

TargetSamplesDistance-W2↓\mathcal{W}_2\downarrowEnergy-W2↓\mathcal{W}_2\downarrowGeometric-W2↓\mathcal{W}_2\downarrow
GMM16Raw0.130 / 0.0391.153 / 0.930NA
Corrected0.005 / 0.0180.040 / 0.039NA
Buffer0.041 / 0.0110.048 / 0.037NA
GMM50Raw0.166 / 0.1324.304 / 4.850NA
Corrected0.048 / 0.1290.321 / 1.340NA
Buffer0.208 / 0.1241.025 / 0.760NA
DW4Raw0.227 / 0.0890.924 / 1.5900.523 / 0.318
Corrected0.192 / 0.1980.109 / 0.1130.482 / 0.496
Buffer0.194 / 0.1890.105 / 0.0990.479 / 0.467
LJ13Raw0.041 / 0.104∗\ast / ∗\ast1.499 / 1.501
Corrected0.026 / 0.0871.405 / 4.0201.626 / 1.970
Buffer0.006 / 0.0050.324 / 0.2501.465 / 1.459

Normalized path-space ESS of the trained sampler at β=1\beta = 1 (SSS / Direct, higher is better).

GMM16GMM50DW4LJ13
0.423 / 0.3860.0077 / 0.00180.46 / 0.500.038 / 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 πβk\pi_{\beta_k}-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.

AuthorsOpenReview ↗

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-W2\mathcal{W}_2 is 0.398±0.1250.398 \pm 0.125, compared with 0.166±0.0340.166 \pm 0.034 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.

Reviewer_Mh6fOpenReview ↗
Rating
4
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposed a neural sampler training pipeline with amortized tempered density training using a diffusion sampler with learnt score function using the Tweedie equation. The training process includes initialization of the approximately exact samples from all temperature levels (using adaptive-SMC), resampling and getting convergence with MCMC finalizing steps, buffering and reusing self-generated importance weighted samples across temperatures. Using architectures that reflect the equivariances in the energy function the sampler is a problem-specific approach to solve complex sampling problems without needing computationally extensive sequential methods.

The setup is reported to surpass recently proposed neural sampling methods. Bringing together diffusion models and the other engineering aspects of sampling is an interesting combination.

Critical strengths

The formulation of Neural sampling as a denoising diffusion problem over multiple temperature levels without having exact samples from the target distribution appears to be novel idea. The combination of this method with resampling, MCMC correction, replay buffers and iterative training has lead to significant improvements over several benchmarks with different complexity levels in comparison to other well-performing and grounded sampling approaches.

The use of replay buffer and generative modeling approach eliminates the need to backpropagate through the sample trajectories, the resampling and MCMC chain which improves the training wall-clock considerably. Although employing an initial sampler and MCMC chains in the training process, this property appears to compensate and improve the training considerably.

Critical weaknesses

The specific contributions of the paper are not listed clearly. To the best of my knowledge most of the parts of the sampling pipeline, the replay buffers and reweighted sample usage, finalizing MCMC steps have been previously introduced in the sampling research (last one in diffusion models). This paper's contribution could be use of the SSS model for sampling, reformulating sampling from multiple temperature levels as a denoising diffusion problem, inputing the temperature or use the specific choices of architectures for each problem or combining multiple existing approaches together. It should be clearly stated in the paper.

Multiple existing pieces require considerable engineering to make the entire sampling pipeline work efficiently. Can the authors discuss this in the paper? what trade offs arise in setting the strength of the initial sampler and the high importance weight variance and the MCMC chain length.

what is the usual support condition (line 130)?

Questions for authors

While equation (7) is consistent and the importance weights are expected to satisfy eq (2), the weight variance is highly dependent on how well the forward diffusion transition steps capture the backward OU discretization. This in turn defines the resampling diversity and required length of the MCMC chain for sample correction. How does this interact with the initalization complexity or the final required MCMC chain?

Can the authors also report the numbers for the initial sampler similar to table 1 for comparison? To see how the training process improved over the original samples.

What portion of the wall-clock time is related to the MCMC chain and which part is related to the diffusion model? As the authors mention this method is useful when evaluation of the target density is expensive. Does that make the MCMC portion considerable?

How do you compare your work to Adjoint sampler "https://arxiv.org/abs/2504.11713", they optimize a sampler "score function" (it's more adjoint approximator) by regressing the endpoint expected gradient energy, which is effectively routed in an equation similar to (4) in the final simplifications.

6 comments
AuthorsOpenReview ↗

Response to Reviewer Mh6f

We would like to thank you for your overall positive assessment of the paper and your detailed comments. This has helped us to revise the manuscript accordingly and to improve its narrative and coverage. We now state the general training procedure and its diffusion instantiation explicitly, compare trained outputs with the initialization, and include an HMC-augmented Adjoint Sampling baseline.

Contribution and implementation

Contribution

The contribution is a general corrected self-training procedure for learning one reusable sampler across a temperature ladder. Each intermediate temperature serves both as a target for the shared proposal and as a source of corrected training data for the next temperature. Proposals are importance-corrected and rejuvenated where they were generated, then reweighted to refresh the adjacent buffer. The shared model is fitted jointly to these repeatedly refreshed populations. The procedure accommodates different proposal families, fitting losses, initialization samplers and invariant MCMC kernels, provided suitable importance weights are available. For example, a normalizing flow can use weighted maximum likelihood, while a diffusion can use score regression.

SSS instantiates this procedure with a diffusion: one network, conditioned on diffusion time and inverse temperature, learns the relative scores across the ladder. Its discrete transitions provide computable path-space importance weights without an auxiliary likelihood model. Training differentiates only the fitting objective, not the sampling, weighting, resampling or MCMC operations. Separate evaluations of raw proposals, importance-corrected proposals and the final buffer distinguish the generator's accuracy from the quality of the corrected and rejuvenated populations. The architectures are standard; the training procedure applies across them.

We believe one important aspect of our contribution is a simple method that works. Our point is not to introduce fancy change-of-measure formulas or novel neural network architectures. Instead, we've found a suitable workflow which is simple to implement, works reasonably well, and is more stable than some other methods, particularly on the limited hardware that we have.

The revised Discussion and Limitation section now states these contributions explicitly.

Practical tradeoffs

Initialization. One adaptive SMC pass selects the temperature ladder, seeds the replay buffers and tunes a local kernel at each temperature. Training reuses these choices. The added ablations examine the SMC particle count, the MCMC kernel and whether to keep D1\mathcal{D}_1 fixed.

Adequate mode coverage matters, but the initial mode weights need not be exact. More initialization effort costs target evaluations upfront and can preserve modes that later local moves rarely reach. Repeated importance correction can improve inaccurate proportions when the proposals cover the relevant regions. It does not guarantee recovery of a missing mode that local MCMC cannot reach. The particle-count ablation and matched-budget SMC comparison examine these effects; the fixed-data amortisation comparison tests what is gained by refreshing the training populations.

Importance-weight variance. Each outer iteration resamples and rejuvenates the current proposals; path weights are not multiplied across outer iterations. We use smaller time steps near the endpoint to limit variability within each path. On LJ13 and Alanine Dipeptide, low path ESS can leave few distinct resampled endpoints, and the buffer gives lower energy discrepancies than fresh importance-corrected samples.

Rejuvenation length. Longer MCMC chains diversify duplicated endpoints at the cost of additional energy and gradient evaluations. On LJ13, 20 / 100 / 2000 rejuvenation moves give:

Number of movesSSS (corrected) Energy-W2\mathcal{W}_2Buffer Energy-W2\mathcal{W}_2
200.742.31
1000.400.37
20000.360.27

Most of the observed reduction occurs between 20 and 100 moves. The revised appendix discusses these tradeoffs.

Support condition

The "usual support condition" at line 130 means that the target is absolutely continuous with respect to the proposal, πβk≪Qθβk\pi_{\beta_k}\ll Q_\theta^{\beta_k}. The revised text states this explicitly: qθ(x∣βk)>0q_\theta(x\mid\beta_k)>0 for πβk\pi_{\beta_k}-almost every xx.

AuthorsOpenReview ↗

Answer from Authors to the Question 1 from Reviewer

Question 1. Weight variance, initialization and MCMC

The weights are density ratios for the implemented discrete sampler, so they remain valid with an imperfect score or a coarse time grid. The proposal and grid affect their variance and hence resampling diversity. For a batch of NN paths with normalized ESS ee, the effective weighted sample size is NeNe; concentrated weights lead to duplicated endpoints after resampling.

Initialization. The SMC ladder controls bridge-weight concentration through its ESS criterion for each Δβ\Delta\beta. Path weights also depend on the learned proposal. Early in training, normalized path ESS can be near its minimum, 1/N1/N. The buffers retain SMC samples while fresh batches gradually replace them through first-in, first-out updates, so training does not rely entirely on early model proposals.

MCMC. Rejuvenation spreads duplicated endpoints within their local basins. We use 1010 moves on GMM16 and DW4, 2020 on GMM50, and 20002000 on LJ13. Longer chains can improve local diversity, but may do little to restore the relative masses of distant modes. MCMC accounts for 84–99.9% of target evaluations, so reducing path-weight variance can reduce the need for costly rejuvenation.

AuthorsOpenReview ↗

Answer from Authors to the Question 2 from Reviewer

Question 2. Initial sampler quality

Before training, the final buffer contains the β=1\beta = 1 population from the initializing SMC sweep. We evaluated it using the same metrics, reference samples and seeds as Table 1. The comparison is included in the revised appendix under Initial sampler. Entries are mean ±\pm standard deviation over 3 seeds, rounded to two decimal places; lower is better. ∗\ast denotes a mean error above 10410^4.

TargetMethodDistance-W2↓\mathcal{W}_2 \downarrowEnergy-W2↓\mathcal{W}_2 \downarrowGeometric-W2↓\mathcal{W}_2 \downarrow
GMM16 (d=16d=16)SMC (initial)0.37±0.150.37 \pm 0.150.61±0.410.61 \pm 0.41—
SSS (raw)0.13±0.090.13 \pm 0.091.15±0.121.15 \pm 0.12—
SSS (corrected)0.01±0.01\mathbf{0.01 \pm 0.01}0.04±0.00\mathbf{0.04 \pm 0.00}—
SSS (buffer)0.04±0.01‾\underline{0.04 \pm 0.01}0.05±0.02‾\underline{0.05 \pm 0.02}—
GMM50 (d=50d=50)SMC (initial)0.57±0.120.57 \pm 0.123.93±1.233.93 \pm 1.23—
SSS (raw)0.17±0.03‾\underline{0.17 \pm 0.03}4.30±0.814.30 \pm 0.81—
SSS (corrected)0.05±0.02\mathbf{0.05 \pm 0.02}0.32±0.06\mathbf{0.32 \pm 0.06}—
SSS (buffer)0.21±0.040.21 \pm 0.041.03±0.20‾\underline{1.03 \pm 0.20}—
DW-4 (d=8d=8)SMC (initial)0.11±0.12\mathbf{0.11 \pm 0.12}0.17±0.020.17 \pm 0.020.38±0.19\mathbf{0.38 \pm 0.19}
SSS (raw)0.23±0.070.23 \pm 0.070.92±0.310.92 \pm 0.310.52±0.140.52 \pm 0.14
SSS (corrected)0.19±0.01‾\underline{0.19 \pm 0.01}0.11±0.03‾\underline{0.11 \pm 0.03}0.48±0.010.48 \pm 0.01
SSS (buffer)0.19±0.010.19 \pm 0.010.11±0.01\mathbf{0.11 \pm 0.01}0.48±0.02‾\underline{0.48 \pm 0.02}
LJ-13 (d=39d=39)SMC (initial)0.00±0.00\mathbf{0.00 \pm 0.00}0.23±0.06\mathbf{0.23 \pm 0.06}1.46±0.00\mathbf{1.46 \pm 0.00}
SSS (raw)0.04±0.030.04 \pm 0.03∗\ast1.50±0.041.50 \pm 0.04
SSS (corrected)0.03±0.030.03 \pm 0.031.41±1.731.41 \pm 1.731.63±0.061.63 \pm 0.06
SSS (buffer)0.01±0.00‾\underline{0.01 \pm 0.00}0.32±0.06‾\underline{0.32 \pm 0.06}1.47±0.00‾\underline{1.47 \pm 0.00}

Using the unrounded discrepancies, corrected SSS reduces the initial Distance-W2\mathcal{W}_2 on the GMMs by about 12–70×12–70 \times and Energy-W2\mathcal{W}_2 by about 12–15×12–15 \times. On DW4, corrected SSS has lower energy error and less variation across seeds (Distance-W2\mathcal{W}_2 standard deviation 0.01 vs 0.12), while the initial SMC population has lower mean distance and geometric errors. On LJ13, initialization is already accurate: the final buffer comes close, but training does not improve accuracy at β=1\beta = 1. The model provides a reusable, β\beta-conditioned proposal; the revised manuscript states both results.

Reviewer_SD4qOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces a new training scheme for amortised samplers. The method tries to learn a temperature-conditioned neural sampler, qθ(x∣βk)q_\theta(x \mid \beta_k) such that qθ(x∣βk)≈πβk(x)q_\theta(x \mid \beta_k) \approx \pi_{\beta_k}(x) at each βk\beta_k (0=β0<⋯<βK=10=\beta_0<\cdots<\beta_K=1). The authors propose the corrected self-training scheme, where the replay buffer Dk\mathcal{D}_k keeps a set of approximate samples of πβk\pi_{\beta_k}. Those approximate samples at each level are from the proposal with the previous temperature βk−1\beta_{k-1}, followed by resampling, MCMC rejuvenation (using preconditioned HMC), and reweighting. The authors then propose an instantiation, sequential score sampler (SSS), where the temperature-conditioned proposal qθ(⋅∣β)q_\theta(\cdot \mid \beta) follows a reverse-time diffusion process with the score model trained by target-score matching. For SSS, because marginal densities are not available, importance weights are computed in the path space. The proposed methods were evaluated on synthetic targets (GMMs and DW-4) and the Lennard-Jones potential (LJ-13) against two temperature-annealed training methods (TA-BG and PITA).

Critical strengths
  1. The writing is very easy to follow, which is especially important right now, as tons of LLM slop are being submitted to the venue.
  2. The idea of the global proposal within SMC by a shared network sounds reasonable, especially if one can incorporate a strong inductive bias into the neural network architecture.
Critical weaknesses

1. Results are not convincing

Although the proposed method (SSS) shows promising results compared to baselines, I'm not fully convinced by the evaluation.

(1) Regarding the baselines, I think the authors should include SMC + preconditioned HMC rejuvenation with at least the same budget for a single outer loop. If SMC alone already solves the problem, then amortisation can be done simply by reverse KL training using the SMC samples.

(2) While shorter wall-clock time is good to have, I believe researchers in this field care more about the number of energy evaluations. The authors should report the number of energy (and gradient) evaluations for all the algorithms considered.

(3) The benchmark suite is limited. Both PITA and TA-BG consider alanine dipeptide.

2. Discussion is not enough

The paper lacks a conclusion and limitations section. I think adding them at the end of the paper would strengthen it. Especially, I think the main limitation is that the proposed method can accumulate error. If Dk\mathcal{D}_k does not correctly approximate πβk\pi_{\beta_k}, e.g., misses a mode that can't be reached from other modes through short MCMC rejuvenation, then self-training on the samples would never correct it. There is only a short discussion of this in the last paragraph of Section 3.1, but it should be discussed more thoroughly in a dedicated limitations section. The bias from self-normalization is also worth mentioning.

Questions for authors
  1. I wonder if the comparison to the baseline is done fairly. (1) The ablation study shows that MCMC rejuvenation is essential. I wonder whether you used the same MCMC for PITA initialisation. (2) As far as I could tell, all the baselines have different network architectures. Are the network sizes matched across the architectures?
  2. I want to see more ablations. (1) You kept D1\mathcal{D}_1 fixed. How critical is that? (2) I found that the number of particles in SMC is quite large. How sensitive is the method to the number of particles? (3) How sensitive SSS is to different MCMC kernels (e.g., MALA, random-walk Metropolis-Hastings)?
9 comments
AuthorsOpenReview ↗

Response to Reviewer SD4q

We would like to thank you for your insightful comments which have helped us to significantly improve the manuscript.

We have added matched-budget SMC comparisons, a fixed-data amortisation baseline, energy and gradient evaluation counts, and the requested ablations. We also extended the benchmarks to Alanine Dipeptide and added a discussion of limitations.

Experimental comparisons

SMC with matched budgets

We compare SSS with a single SMC run using the budget of the entire SSS run, including initialization and all outer iterations. SMC uses the same number of particles, HMC kernel and tempering rule as the SSS warm-up, with more HMC moves at each temperature to match the total cost. We match cost in two ways:

  • SMC with same MCMC budget. SMC performs as many HMC moves in total as SSS does, counting both its warm-up and its rejuvenation steps.
  • SMC with same energy-evaluation budget. SMC calls the energy and its gradient as many times as SSS does in total.

The tables report mean ±\pm standard deviation over three seeds; lower is better. mm is the number of HMC moves per temperature that matches the stated budget, and ∗\ast denotes a mean error above 10410^4.

Gaussian mixtures

TargetSample setDistance-W2↓\mathcal{W}_2 \downarrowEnergy-W2↓\mathcal{W}_2 \downarrow
GMM16SMC, same MCMC budget (m=35m=35)0.194±0.0690.194 \pm 0.0690.274±0.1240.274 \pm 0.124
SMC, same energy budget (m=40m=40)0.141±0.0380.141 \pm 0.0380.188±0.0620.188 \pm 0.062
SSS (buffer)0.041±0.0090.041 \pm 0.0090.048±0.0160.048 \pm 0.016
SSS (corrected)0.005±0.0050.005 \pm 0.0050.040±0.0040.040 \pm 0.004
SSS (raw)0.130±0.0910.130 \pm 0.0911.153±0.1211.153 \pm 0.121
GMM50SMC, same MCMC and energy budget (m=120m=120)0.383±0.1280.383 \pm 0.1282.240±1.0202.240 \pm 1.020
SSS (buffer)0.208±0.0370.208 \pm 0.0371.025±0.2011.025 \pm 0.201
SSS (corrected)0.048±0.0220.048 \pm 0.0220.321±0.0590.321 \pm 0.059
SSS (raw)0.166±0.0340.166 \pm 0.0344.304±0.8074.304 \pm 0.807

Particle systems

TargetSample setDistance-W2Energy-W2Geometric-W2
DW4SMC, same MCMC budget (m=40m=40)0.219±0.0070.219 \pm 0.0070.151±0.0240.151 \pm 0.0240.568±0.0260.568 \pm 0.026
SMC, same energy budget (m=45m=45)0.081±0.0330.081 \pm 0.0330.124±0.0240.124 \pm 0.0240.312±0.0410.312 \pm 0.041
SSS (buffer)0.194±0.0100.194 \pm 0.0100.105±0.0140.105 \pm 0.0140.479±0.0160.479 \pm 0.016
SSS (corrected)0.192±0.0100.192 \pm 0.0100.109±0.0340.109 \pm 0.0340.482±0.0140.482 \pm 0.014
SSS (raw)0.227±0.0750.227 \pm 0.0750.924±0.3130.924 \pm 0.3130.523±0.1440.523 \pm 0.144
LJ13SMC, same MCMC budget (m=17,343m=17{,}343)0.006±0.0010.006 \pm 0.0010.251±0.0280.251 \pm 0.0281.458±0.0011.458 \pm 0.001
SMC, same energy budget (m=25,927m=25{,}927)0.005±0.0010.005 \pm 0.0010.262±0.0440.262 \pm 0.0441.461±0.0011.461 \pm 0.001
SSS (buffer)0.006±0.0020.006 \pm 0.0020.324±0.0630.324 \pm 0.0631.465±0.0031.465 \pm 0.003
SSS (corrected)0.026±0.0270.026 \pm 0.0271.405±1.7331.405 \pm 1.7331.626±0.0591.626 \pm 0.059
SSS (raw)0.041±0.0320.041 \pm 0.032∗\ast1.499±0.0431.499 \pm 0.043

On the GMM targets, the SSS buffer reduces mean errors relative to SMC by factors of 3–6 on GMM16 and about 2 on GMM50, with less variation across seeds. The residual SMC errors are consistent with inaccurate mode weights. Along the tempering path, the modes separate while the lower-weight component's mass still has to fall from about 60% to 20%. Once the barrier is high, additional local HMC moves do little to adjust these proportions; SMC relies on reweighting and resampling, reducing particle diversity. SSS draws fresh proposals at every iteration and corrects their contributions with importance weights. SSS (corrected) has lower mean error than the buffer on all four GMM metrics. On GMM16, distance error falls from 0.041 to 0.005 and energy error from 0.048 to 0.040.

On DW4, the SSS buffer has lower mean errors on all three metrics at the same MCMC budget. At the same energy budget, the buffer has lower energy error, while SMC has lower distance and geometric errors. The buffer has smaller standard deviations on all three metrics than energy-matched SMC; against MCMC-matched SMC, its energy and geometric standard deviations are smaller, and its distance standard deviation is larger.

On LJ13, energy-matched SMC has lower mean errors than the SSS buffer on all three metrics. MCMC-matched SMC matches the buffer's distance error at the displayed precision (0.006) and has lower energy and geometric errors.

AuthorsOpenReview ↗

Response to Reviewer SD4q

Amortising fixed SMC samples

To test amortisation of fixed SMC samples, we fill the buffers with the mixture populations from matched energy-evaluation budget reported above and fit the same β\beta-conditioned diffusion. We retain the loss, optimiser and numbers of outer and inner iterations of SSS, but disable buffer refreshes: the backward SDE, importance weighting, HMC rejuvenation and bridging to the next temperature. The training populations remain fixed.

The target score-matching loss uses the energy gradients stored with the samples. Computing these gradients adds at most about 1% to the SMC energy-evaluation budget, so the total target-evaluation cost is at least that of SSS.

To compare the learned samplers directly, we evaluate raw backward-SDE samples from both models, without importance correction.

TargetSamplerDistance-W2↓\mathcal{W}_2 \downarrowEnergy-W2↓\mathcal{W}_2 \downarrow
GMM16SMC + amortisation0.188±0.0420.188 \pm 0.0421.166±0.1061.166 \pm 0.106
SSS (raw)0.130±0.0910.130 \pm 0.0911.153±0.1211.153 \pm 0.121
GMM50SMC + amortisation0.398±0.1250.398 \pm 0.1255.462±0.8735.462 \pm 0.873
SSS (raw)0.166±0.0340.166 \pm 0.0344.304±0.8074.304 \pm 0.807

On GMM50, SSS reduces mean Distance-W2\mathcal{W}_2 by a factor of 2.4 and mean Energy-W2\mathcal{W}_2 by a factor of 1.3. On GMM16, SSS has the lower mean distance error, and the two are comparable in energy.

On GMM50, the fitted sampler's Distance-W2\mathcal{W}_2 follows that of its SMC training set seed by seed: 0.28, 0.39 and 0.53 for the trained sampler, against 0.22, 0.40 and 0.53 for the SMC samples. In this experiment, fitting the fixed populations leaves their errors largely intact. SSS regenerates training data from the current model at every outer iteration and corrects it with importance weights and HMC under the true energy. Its buffer errors are smaller by factors of roughly 22–44 than those of the energy-matched SMC populations on the mixtures.

The fixed-data amortisation experiment concerns only the mixtures. Comparisons with PITA and TA-BG are described in Question 1 below.

Energy and gradient evaluation budgets

The following table reports total energy and energy-gradient evaluations, in millions:

TargetMethodEnergy evals (M)Gradient evals (M)
GMM16SSS (ours)85.741.7
PITA411.6368.4
TA-BG86.944.9
ASMatched SSSMatched SSS
GMM50SSS (ours)230.5114.1
PITA551.2438.6
TA-BG227.0115.3
ASMatched SSSMatched SSS
DW4SSS (ours)50.224.5
PITA374.6350.7
TA-BG98.073.8
ASMatched SSSMatched SSS
LJ13SSS (ours)9,0773,038
ASMatched SSSMatched SSS
ALDPSSS (ours)420.2208.1
ASMatched SSSMatched SSS

The counts are averaged over 3 seeds. They vary across seeds because SMC can select different numbers of temperatures and kernel tuning can require different numbers of bisection steps. For LJ13, our PITA and TA-BG runs diverged, so we use their published results from the PITA paper.

Alanine Dipeptide and Adjoint Sampling

We added Alanine Dipeptide, which is also evaluated by PITA and TA-BG. We also added Adjoint Sampling, which shares endpoint-gradient regression with SSS and includes DW4 and LJ13 among its original benchmarks. Table 2 of the general response reports the Alanine Dipeptide results; the AS comparison settings are described in Question 1 below.

Discussion and limitations

The revised discussion addresses missing modes, propagation of errors through self-training, and finite-sample bias from self-normalized importance sampling. Importance weighting cannot recover a mode absent from the proposed samples, and short local MCMC runs may fail to reach it. We also expanded the discussion of related work.

AuthorsOpenReview ↗

Answer from Authors to the Question 1 from Reviewer

Question 1: Baseline settings

The original submission included MCMC moves for the baselines; the new runs match their nominal HMC budgets to SSS.

We reran PITA and TA-BG on GMM16, GMM50, and DW4 using the same positive inverse temperatures selected by SSS's SMC initialization for each seed. For PITA on GMM16 and GMM50, each of its two networks uses exactly the same ResMLP architecture as SSS. For DW4, we retain PITA's native EGNN architectures and adjust their widths so that each network matches SSS's parameter count within 0.8%. PITA therefore retains approximately twice SSS's total parameter count, reflecting its two-network formulation. For TA-BG, we retain its spline-flow architecture and adjust the conditioner widths so that the entire flow matches SSS's single network within 0.9%.

Each baseline receives the same nominal number of HMC particle proposals as SSS's SMC initialization and training rejuvenation combined. We use the same HMC kernel and distribute the allowance across annealing transitions. PITA applies HMC after each transition; TA-BG applies it to the importance-resampled particles before fitting the flow to the next training buffer.

The table reports mean ±\pm standard deviation over 3 seeds for the rerun baselines; lower is better.

TargetMethodDistance-W2↓\mathcal{W}_2\downarrowEnergy-W2↓\mathcal{W}_2\downarrowGeometric-W2↓\mathcal{W}_2\downarrow
GMM16 (d=16d=16)TA-BG0.023±0.0050.023 \pm 0.0050.067±0.0430.067 \pm 0.043NA
PITA0.168±0.0300.168 \pm 0.0300.984±0.0820.984 \pm 0.082NA
GMM50 (d=50d=50)TA-BG0.114±0.1530.114 \pm 0.1531.065±1.6401.065 \pm 1.640NA
PITA2.115±0.0212.115 \pm 0.021240.014±1.016240.014 \pm 1.016NA
DW4 (d=8d=8)TA-BG0.317±0.0160.317 \pm 0.01624.519±10.46324.519 \pm 10.4630.789±0.0220.789 \pm 0.022
PITA0.150±0.1240.150 \pm 0.1242.664±2.1072.664 \pm 2.1070.479±0.2000.479 \pm 0.200

We implement AS + HMC in the same code base as SSS. AS regresses on an endpoint target built from ∇E\nabla E at its own samples, which it retains in a uniformly weighted FIFO buffer. Our variant initializes this buffer from SSS's SMC population at β=1\beta = 1 for the same seed and rejuvenates each new batch at β=1\beta = 1 before insertion. It uses the HMC kernel adapted by SMC at β=1\beta = 1 and the same number of moves per sample as SSS training. SSS itself rejuvenates at β1,…,βM−1\beta_1, \ldots, \beta_{M−1} and reaches β=1\beta = 1 through bridge weights. AS + HMC trains at β=1\beta = 1 alone, without a temperature ladder, path-space importance resampling or bridge weights.

The original AS uses a memoryless VE process; we run its loop on SSS's OU process, where its endpoint target coincides with SSS's TSM target at β=1\beta = 1. The network, time grid and optimizer are identical, so parameter counts match exactly. For each seed, we also match the total number of backward-SDE samples, gradient steps and total mini-batch size; the AS buffer has the capacity of one SSS level. Including SMC initialization, AS + HMC uses 0.005–2% fewer energy-gradient evaluations than SSS. We run it on GMM16, GMM50, DW4, LJ13, and ALDP with three seeds each.

The general response and revised submission report the full comparison with all three SSS sample sets.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General response to Reviewers

We would like to thank the reviewers for their detailed comments.

We have revised the manuscript to clarify the contribution and extend the experimental evaluation. We have highlighted in red{\color{red}\text{red}} the edited parts in our revised submission compared to our original submission.

Contribution

We introduce a general corrected self-training procedure for learning a temperature-conditioned sampler. The procedure separates the construction of corrected training populations from the proposal family and its fitting objective. Current proposals are importance-corrected and rejuvenated at intermediate temperatures, then reweighted to refresh training buffers at adjacent temperatures. One shared model learns jointly from these populations. The proposal can, for example, be a normalizing flow fitted by weighted maximum likelihood or a diffusion fitted by score regression, provided suitable importance weights are available. SSS is the diffusion instantiation, with computable weights for its discrete sampling paths and no auxiliary likelihood model.

We believe one important aspect of our contribution is a simple method that works. Our point is not to introduce fancy change-of-measure formulas or novel neural network architectures. Instead, we've found a suitable workflow which is simple to implement, works reasonably well, and is more stable than some other methods, particularly on the limited hardware that we have.

Learning beyond the initial samples

We added comparisons with SMC using the energy-evaluation budget of the entire SSS run, including initialization and training. On GMM16, the SSS buffer reduces mean distance discrepancy from 0.1410.141 to 0.0410.041; on GMM50, it reduces mean energy discrepancy from 2.2402.240 to 1.0251.025. On DW4, SMC has lower distance and geometric discrepancies, while the SSS buffer has a slightly lower energy discrepancy. On LJ13, SMC has lower means on all three metrics.

We also fitted the same temperature-conditioned diffusion to fixed SMC populations, retaining the loss and training schedule of SSS. On GMM50, refreshing the training data through SSS reduces the raw sampler's mean distance discrepancy from 0.3980.398 to 0.1660.166. The difference is smaller on GMM16, where the energy discrepancies are similar. These comparisons, reported under Additional baselines, test the benefit of corrected self-training beyond fitting a fixed Monte Carlo dataset.

Baselines, evaluation costs, and molecular sampling

We reran PITA and TA-BG on GMM16, GMM50, and DW4 with matched temperature ladders, closely matched network sizes, and the same nominal HMC proposal budget. We added an HMC-augmented AS reimplementation using SSS's architecture, initialization, and matched computational settings, and evaluated Alanine Dipeptide as an additional target. Energy and energy-gradient counts are reported under Computational cost. LJ13 and Alanine Dipeptide results for PITA and TA-BG are taken from the published PITA benchmark; those entries are not locally budget-matched comparisons.

Sample sets and limitations

We report raw proposals, importance-corrected proposals, and final-buffer samples separately. On the mixtures, correction gives the most accurate SSS samples in the reported distance and energy metrics. On Alanine Dipeptide, raw samples give the lowest mean angular discrepancy, while the buffer gives the lowest energy discrepancies among the compared methods. Buffer draws reuse a finite stored population; fresh model proposals can be generated after training.

The revised discussion addresses missing modes, finite-sample bias from self-normalization, and concentrated importance weights. Additional ablations examine the fixed first buffer, the initialization particle count, and the MCMC kernel. In the eccentric-mixture example, training improves the weight of an underrepresented mode, but does not recover a mode lost during initialization.

AuthorsOpenReview ↗

Benchmark results

The results are produced in three random seeds. AS + HMC denotes our augmented reimplementation; the individual responses and revised manuscript describe its settings.

Table 1. Sampler performance on GMM16, GMM50, DW4, and LJ13. Entries are mean ±\pm standard deviation. For each target and metric, bold marks the lowest mean at the displayed precision and underlining marks the second-lowest distinct mean. Equal displayed means share the same formatting. A dash indicates that the metric does not apply. ∗\ast denotes a mean above 10410^4. LJ13 results for TA-BG and PITA retain the precision of the published source.

TargetMethodDistance-W2↓\mathcal{W}_2 \downarrowEnergy-W2↓\mathcal{W}_2 \downarrowGeometric-W2↓\mathcal{W}_2 \downarrow
GMM16 (d=16d=16)TA-BG0.023±0.005‾\underline{0.023 \pm 0.005}0.067±0.0430.067 \pm 0.043—
PITA0.168±0.0300.168 \pm 0.0300.984±0.0820.984 \pm 0.082—
AS + HMC2.221±0.0722.221 \pm 0.0729.947±0.3799.947 \pm 0.379—
SSS (raw)0.130±0.0910.130 \pm 0.0911.153±0.1211.153 \pm 0.121—
SSS (corrected)0.005±0.005\mathbf{0.005 \pm 0.005}0.040±0.004\mathbf{0.040 \pm 0.004}—
SSS (buffer)0.041±0.0090.041 \pm 0.0090.048±0.016‾\underline{0.048 \pm 0.016}—
GMM50 (d=50d=50)TA-BG0.114±0.153‾\underline{0.114 \pm 0.153}1.065±1.6401.065 \pm 1.640—
PITA2.115±0.0212.115 \pm 0.021240.014±1.016240.014 \pm 1.016—
AS + HMC2.338±0.0602.338 \pm 0.06032.965±1.29432.965 \pm 1.294—
SSS (raw)0.166±0.0340.166 \pm 0.0344.304±0.8074.304 \pm 0.807—
SSS (corrected)0.048±0.022\mathbf{0.048 \pm 0.022}0.321±0.059\mathbf{0.321 \pm 0.059}—
SSS (buffer)0.208±0.0370.208 \pm 0.0371.025±0.201‾\underline{1.025 \pm 0.201}—
DW4 (d=8d=8)TA-BG0.317±0.0160.317 \pm 0.01624.519±10.46324.519 \pm 10.4630.789±0.0220.789 \pm 0.022
PITA0.150±0.124\mathbf{0.150 \pm 0.124}2.664±2.1072.664 \pm 2.1070.479±0.200\mathbf{0.479 \pm 0.200}
AS + HMC0.467±0.0900.467 \pm 0.0900.914±0.2840.914 \pm 0.2840.956±0.1260.956 \pm 0.126
SSS (raw)0.227±0.0750.227 \pm 0.0750.924±0.3130.924 \pm 0.3130.523±0.1440.523 \pm 0.144
SSS (corrected)0.192±0.010‾\underline{0.192 \pm 0.010}0.109±0.034‾\underline{0.109 \pm 0.034}0.482±0.014‾\underline{0.482 \pm 0.014}
SSS (buffer)0.194±0.0100.194 \pm 0.0100.105±0.014\mathbf{0.105 \pm 0.014}0.479±0.016\mathbf{0.479 \pm 0.016}
LJ13 (d=39d=39)TA-BG1.21±0.021.21 \pm 0.0261.47±0.1261.47 \pm 0.124.16±0.014.16 \pm 0.01
PITA0.04±0.000.04 \pm 0.002.26±0.212.26 \pm 0.211.65±0.001.65 \pm 0.00
AS + HMC0.899±0.7700.899 \pm 0.770∗\ast3.179±1.4853.179 \pm 1.485
SSS (raw)0.041±0.0320.041 \pm 0.032∗\ast1.499±0.043‾\underline{1.499 \pm 0.043}
SSS (corrected)0.026±0.027‾\underline{0.026 \pm 0.027}1.405±1.733‾\underline{1.405 \pm 1.733}1.626±0.0591.626 \pm 0.059
SSS (buffer)0.006±0.002\mathbf{0.006 \pm 0.002}0.324±0.063\mathbf{0.324 \pm 0.063}1.465±0.003\mathbf{1.465 \pm 0.003}

Table 2. Sampler performance on Alanine Dipeptide. Discrepancies and ranks follow Table 1. PITA and TA-BG results are from the published PITA benchmark. SSS and AS + HMC each used approximately 2×1082\times10^8 energy evaluations. We use the same metrics as the revised manuscript, which explains the omission of Rama-KL.

MethodEnergy-W1↓\mathcal{W}_1 \downarrowEnergy-W2↓\mathcal{W}_2 \downarrowT-W2↓\mathcal{W}_2 \downarrow
TA-BG83.457±0.07083.457 \pm 0.07086.176±0.10486.176 \pm 0.1040.979±0.0120.979 \pm 0.012
PITA1.530±0.0681.530 \pm 0.0681.615±0.0531.615 \pm 0.0530.270±0.023‾\underline{0.270 \pm 0.023}
AS + HMC∗\ast∗\ast1.490±0.0031.490 \pm 0.003
SSS (raw)3.765±0.3023.765 \pm 0.3023.858±0.3053.858 \pm 0.3050.222±0.058\mathbf{0.222 \pm 0.058}
SSS (corrected)1.230±1.422‾\underline{1.230 \pm 1.422}1.432±1.525‾\underline{1.432 \pm 1.525}0.757±0.7450.757 \pm 0.745
SSS (buffer)0.212±0.190\mathbf{0.212 \pm 0.190}0.257±0.181\mathbf{0.257 \pm 0.181}0.399±0.2070.399 \pm 0.207

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.