acceptodds
Under review as a conference paper at ICLR 2027

ENCORE: Exact Non-equilibrium COntrol with Replica Exchange for Diffusion Generation

Abstract

Inference-time control steers a pretrained generative model towards a target distribution without retraining. We study tilted targets , where is the sampler output distribution and is an evaluable reweighting function. Existing approaches rely on sequential annealing with sequential Monte Carlo (SMC) or parallel annealing with replica exchange (RE). Sequential control is exact but needs large particle populations, whereas no exact parallel control method exists: existing RE corrections approximate an intractable time reversal and are biased. We propose Exact Non-equilibrium COntrol with Replica Exchange (ENCORE), the first exact parallel control method. Each replica stores its generation trajectory, so the upward move is a truncation and the intractable time reversal is never simulated. We prove target invariance and show that the resulting dynamics are those of non-equilibrium replica exchange with the exact time reversal as forward proposal. Under regularity conditions, our diffusion analysis shows that both sequential and parallel control become unstable under refinement of the time discretisation without guidance, whereas guided proposals remain stable and yield diagnostics for tuning the schedule and the computational budget. Across synthetic targets, Boltzmann sampling of biomolecules, and image generation, ENCORE achieves competitive accuracy and diversity, remains robust to sampler perturbations, and applies to distilled samplers where existing RE corrections are unavailable.

Primary area:
generative models
Keywords:
Diffusion models, Inference-time computation

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_45EXOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper addresses inference-time control of a pretrained sampler. Given a sampler with output law p₀ and a non-negative reward or likelihood G₀, the goal is to sample from G₀p₀ without retraining. Specifically, ENCORE has each replica carry its partial denoising trajectory. Moving a replica up the ladder is then a truncation, which is exactly a draw from the true time reversal. The intractable marginals cancel, and the Metropolis–Hastings ratio reduces to the SMC backward weight evaluated at two points. In the unguided case this is a ratio of potentials only. The cost is O(N²d) memory.

Critical strengths
  1. Removes a real bias in the closest prior method. The construction is simple and checkable, and correctness does not depend on the model exactly reversing a noising process.

  2. Broader applicability. The unguided acceptance uses only potential values. It therefore applies to distilled or deterministic samplers with no noising kernel.

Critical weaknesses
  1. Limited algorithmic novelty. Path-space lifting is standard in SMC. Appendix C itself notes that the exchange weight is the SMC incremental weight on the trajectory ladder. The contribution is transplanting this into replica exchange.

  2. Theory–experiment mismatch. The favourable result (finite Λ) concerns the guided variant. However, every experiment whose configuration is stated uses unguided proposals.

  3. No demonstrated advantage over naive reweighting. Plain self-normalised importance sampling (SNIS) beats ENCORE in every column of Table 4 and in the fixed-budget tables. On SDXL-Turbo, best-of-64 nearly matches ImageReward (1.651 vs 1.711) with higher diversity and 100% unique outputs, versus 84% for ENCORE.

Questions for authors
  1. Your main positive result, Thm. 3.2(b), shows that the guided variant has a finite communication barrier. As far as I can tell from App. E, however, all experiments use unguided proposals. Could you report results for guided ENCORE on at least one task?

  2. Storage complexity and scalability to larger data. ENCORE stores the full partial trajectory at every level, so each ladder holds O(N²d) memory. The experiments keep this small: 5 levels for images, and 22–63-atom molecules elsewhere. I am concerned that the method becomes hard to apply to larger state spaces such as high-resolution images, proteins, etc.

  3. Sampling efficiency. Could you report restart rates, round-trip rates, or effective sample size per NFE alongside accuracy? This would show how much of ENCORE’s output is effectively independent compared with SMC baselines.

4 comments
AuthorsOpenReview ↗

Response to Reviewer 45EX [1/3]

We thank the reviewer for their time and feedback on our manuscript. We address each comment below, providing clarification and additional results where appropriate.

Guided ENCORE (Q1, W2)

Could you report results for guided ENCORE on at least one task?

Yes. Tab. 1 below compares guided and unguided ENCORE on eight tasks. Guidance lowered mean error on seven of the eight. In molecular tempering, guided ENCORE achieved lower mean error on all four metrics for all three molecules, under the same physical-energy target and a budget of 4.22M denoiser evaluations per run. Please see Tab. 1 in our response to Reviewer n25t for these results alongside CREPE, TDS and FKD.

Table 1. Sampling accuracy of SNIS and ENCORE.

Experiment (Metric ↓)SNISUnguided ENCOREGuided ENCORE
Gaussian shift (CDF Error)0.9533 ± 0.02020.8729±0.0474‾\underline{0.8729±0.0474}0.0248 ± 0.0108
ALDP 300 K (Energy TVD)0.6289 ± 0.03300.1955±0.0342‾\underline{0.1955±0.0342}0.1689 ± 0.0097
A4 500 K (Energy TVD)0.2715 ± 0.02280.1502±0.0172‾\underline{0.1502±0.0172}0.1418 ± 0.0075
A6 600 K (Energy TVD)0.1793 ± 0.01570.1107±0.0123‾\underline{0.1107±0.0123}0.1067 ± 0.0048
1ENH (RMSD W1\rm W_1)62.14 ± 33.6342.59±6.36‾\underline{42.59±6.36}37.33 ± 12.99
1BTA (RMSD W1\rm W_1)2.32 ± 1.271.98±1.96‾\underline{1.98±1.96}1.36 ± 0.33
1TTG (RMSD W1\rm W_1)63.55 ± 103.0911.29±9.62‾\underline{11.29±9.62}5.48 ± 1.70
2RN2 (RMSD W1\rm W_1)8.24 ± 2.346.95 ± 0.878.20±1.46‾\underline{8.20±1.46}

Entries are mean ± standard deviation over 32 seeds (Gaussian shift), five (molecular tempering) or three (proteins).

Scalability to Larger Data (Q2)

I am concerned that the method becomes hard to apply to larger state spaces such as high-resolution images, proteins, etc.

To show that ENCORE can be applied to larger molecular state spaces, we extend our experiments to four proteins with 48–155 residues and perform enhanced sampling targeting the biased distribution

π(x)∝p0(x)exp⁡[−α FNC(x)],\pi(x) \propto p_0(x)\exp [ -\alpha\ \mathrm{FNC}(x) ],

where p0p_0 is the distribution over protein backbone conformations generated by BioEmu v1.1 using an Euler-discretised sampler, and FNC(x)\mathrm{FNC}(x) is a differentiable fraction of native contacts (FNC) computed from Cα\mathrm{C}_\alpha distances. We set the bias strength to α=8\alpha=8 for all four proteins, favouring conformations with lower FNC.

Each residue contributes three translational and three rotational degrees of freedom, giving nominal sampler-state dimensions of 288, 534, 564 and 930 for 1ENH, 1BTA, 1TTG and 2RN2. With three Cartesian coordinates per atom in the alanine-peptide experiments, the largest protein has about five times the state dimension of our largest peptide system.

Tabs. 2 and 3 below report RMSD-distribution error and backbone geometric validity as mean ± standard deviation over three seeds.

Table 2. RMSD-distribution Wasserstein error (10−310^{-3} nm; lower is better).

Method1ENH1BTA1TTG2RN2
ENCORE-G37.33 ± 12.991.36±0.33‾\underline{1.36±0.33}5.48±1.70‾\underline{5.48±1.70}8.20±1.46‾\underline{8.20±1.46}
ENCORE-U42.59 ± 6.361.98 ± 1.9611.29 ± 9.626.95 ± 0.87
CREPE-G166.85 ± 66.0225.13 ± 12.9518.32 ± 11.7518.49 ± 11.62
CREPE-U125.09 ± 24.5225.72 ± 14.1016.09 ± 4.7510.76 ± 1.19
TDS40.24±12.54‾\underline{40.24±12.54}1.30 ± 0.245.38 ± 1.068.63 ± 2.45
FKD43.70 ± 16.263.03 ± 0.7622.01 ± 14.2412.34 ± 1.22
SNIS62.14 ± 33.632.32 ± 1.2763.55 ± 103.098.24 ± 2.34

Table 3. Backbone geometric validity (%; higher is better).

Method1ENH1BTA1TTG2RN2
ENCORE-G98.19±0.24‾\underline{98.19±0.24}96.11±0.73‾\underline{96.11±0.73}92.64±0.48‾\underline{92.64±0.48}91.94 ± 1.75
ENCORE-U98.20 ± 0.3895.87 ± 0.1192.66 ± 0.4091.56 ± 0.32
CREPE-G71.11 ± 7.7389.44 ± 4.0087.22 ± 6.5483.89 ± 3.59
CREPE-U77.50 ± 5.6890.15 ± 5.1586.97 ± 2.3786.78 ± 0.29
TDS97.55 ± 0.1396.63 ± 0.4992.48 ± 0.6191.76±0.15‾\underline{91.76±0.15}
FKD97.82 ± 0.1595.90 ± 0.4392.09 ± 0.6191.15 ± 0.44
SNIS91.90 ± 10.0295.87 ± 0.4486.61 ± 8.4491.12 ± 0.30

ENCORE performs competitively across the four proteins. ENCORE-G ranks first or second in RMSD-distribution error on every protein, and ENCORE-U ranks first on the largest protein (2RN2) and third on the other three. Both variants maintain high backbone geometric validity and outperform CREPE-G and CREPE-U on both metrics for every protein.

AuthorsOpenReview ↗

Response to Reviewer 45EX [2/3]

Memory (Q2)

ENCORE stores the full partial trajectory at every level, so each ladder holds O(N2d)O(N^2d) memory.

We agree that ENCORE incurs an additional storage cost, as noted in our limitations (Sec. 5). For NN replica levels and dd-dimensional states, storing the paths requires O(N2d)O(N^2d) sampler-state memory, compared with O(Nd)O(Nd) for CREPE and O(Md)O(Md) for SMC methods such as TDS with MM parallel particles. In exchange, the stored paths let ENCORE move up the ladder by truncation, without additional denoiser evaluations.

ENCORE can be preferable when memory or parallelism is fixed and samples are needed online. Its path storage stays constant as samples accumulate. Counting states at replica levels, a single ENCORE ladder stores fewer states than MM parallel SMC particles when M>N2M>N^2. In our molecular tempering experiments, N=15N=15 and each run retains M=49,000M=49{,}000 samples, so N2=225≪MN^2=225\ll M.

Sampling Efficiency (Q3)

Could you report restart rates, round-trip rates, or effective sample size per NFE alongside accuracy?

To assess practical sampling efficiency, we compare unguided (U) and guided (G) ENCORE on the same 15-replica ladder. Each run used one NVIDIA H100 and retained 49,000 outputs after 1,000 warm-up half-sweeps. Tabs. 4–5 below report communication rates and effective sample size (ESS) per denoiser evaluation, as mean ± sample standard deviation over five seeds per configuration.

Table 4. Exchange rejection and communication on the same replica ladder.

TargetSummed rejection U ↓Summed rejection G ↓Restarts/1,000 half-sweeps U ↑Restarts/1,000 half-sweeps G ↑Round trips/h U ↑Round trips/h G ↑
ALDP 300 K6.185 ± 0.0132.282 ± 0.08911.66 ± 0.4288.52 ± 2.9067.66 ± 2.43314.78 ± 10.08
A4 500 K4.289 ± 0.0031.629 ± 0.00834.23 ± 0.49114.11 ± 1.73192.28 ± 5.00387.13 ± 12.14
A6 600 K3.322 ± 0.0171.373 ± 0.00855.21 ± 1.25135.99 ± 0.52298.92 ± 5.43452.33 ± 3.03

Guidance reduces summed rejection and increases both restart frequency and round trips per hour on all three targets.

Table 5. Estimated ESS per million denoiser evaluations. One NFE is one denoiser forward evaluation on a single molecular state.

TargetObservableENCORE-U ESS/million NFEs ↑ENCORE-G ESS/million NFEs ↑Paired G/U ratio ↑
ALDP 300 KEnergy67.1 ± 20.71,136.0 ± 96.418.48× ± 6.71
ALDP 300 KTorsion66.4 ± 44.2575.2 ± 105.813.03× ± 9.55
A4 500 KEnergy187.5 ± 54.9804.2 ± 141.14.63× ± 1.72
A4 500 KTorsion147.0 ± 14.2414.0 ± 55.12.86× ± 0.59
A6 600 KEnergy301.8 ± 65.7999.9 ± 237.93.37× ± 0.73
A6 600 KTorsion237.2 ± 78.9520.3 ± 85.82.46× ± 1.07

We estimate ESS from autocorrelations of each chronological post-warm-up trace, retaining repeated states. Torsion denotes the sine of the first backbone ψ\psi angle. We compute G/U ratios within each seed before averaging.

Across all three targets, guided ENCORE completed more round trips per hour and achieved lower mean error on all four accuracy metrics at matched NFE budgets (see Tab. 1 in our response to Reviewer n25t). Guidance also increased ESS per NFE for both observables in every seed on every target, with mean paired gains of 3.37–18.48× for energy and 2.46–13.03× for torsion.

AuthorsOpenReview ↗

Response to Reviewer 45EX [3/3]

Connection between SMC and RE (W1)

Path-space lifting is standard in SMC.

Appendix C itself notes that the exchange weight is the SMC incremental weight on the trajectory ladder. The contribution is transplanting this into replica exchange.

Path-space lifting is common in SMC, but as far as we know, it has not been used in diffusion and replica exchange before this work. SMC moves in one direction, so path space fits naturally. Replica exchange moves in both directions and is usually used to get marginal samples, not paths. Moving up the ladder needs the sampler's time reversal, which is not practical, so CREPE is only approximate. By storing the trajectory, this move becomes a truncation. This gives the first exact replica-exchange method for inference-time control, together with the communication barrier theory in Thm. 3.2.

Weight matching in SMC is expected, as it is in parallel tempering and annealed SMC. The two approaches organise the same computation differently: SMC runs MM particles in parallel across NN sequential levels, while replica exchange runs NN levels in parallel across MM sequential iterations. We are not claiming to outperform SMC. We provide the exact counterpart on the replica-exchange side, which was not available before.

SNIS and BoN (W3)

Plain self-normalised importance sampling (SNIS) beats ENCORE in every column of Table 4 and in the fixed-budget tables.

We agree and should have stated this more clearly. SNIS works well when the proposal covers the target, but coverage can deteriorate under stronger tilts or in higher dimensions.

To test this in practice, we compared SNIS and ENCORE on eight tasks: a 1D Gaussian shifted by 8σ8\sigma, three peptide tempering tasks, and four BioEmu protein benchmarks biased towards lower native-contact fractions (see our response to Q2 above). Guided ENCORE had lower mean error than SNIS on all eight (see Tab. 1 above).

On SDXL-Turbo, best-of-64 nearly matches ImageReward (1.651 vs 1.711) with higher diversity and 100% unique outputs, versus 84% for ENCORE.

We agree that Best-of-N (BoN) offers a strong reward–diversity trade-off on the image tasks. For each prompt, BoN returns the highest-reward image from each of 32 independent groups of 64 draws, so its outputs are independent and distinct. ENCORE instead records two outputs from each of 16 independent chains; the two coincide when the chain's target-level replica has not changed, lowering uniqueness and diversity scores. At a fixed budget, fewer but longer chains trade diversity for reward: with 8 chains instead of 16, mean ImageReward rises from 1.711 to 1.745, while DINO-Vendi falls from 3.02 to 2.78 (Tab. 13 of the updated manuscript).

Unlike BoN, ENCORE and the Monte Carlo baselines target q0∝exp⁡{r} p0q_0\propto\exp\{r\}\,p_0. However, these metrics cannot assess target fidelity for any method: reward and generic diversity do not measure discrepancy from q0q_0, and no reliable target reference is available. We therefore interpret the image results as showing that ENCORE is competitive and works with distilled samplers, including DMD2, whose native sampler lacks the noising kernel CREPE requires. We will clarify this scope in the manuscript.


Thank you again for your review. We would be happy to clarify anything further during the discussion period.

Thanks for the addtional experiments and clarification from authors. I believe include the new experiments, clarifications about memory usage, and further performance analysis in the manuscript will make this paper better. I will increase my score.

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

This paper studies the inference of generative model with reward. Existing sequential monte carlo method can suffer from weight and ancestral degeneracy while existing replica exchange method (CREPE) requires a tractable paired noising kernel. To address this, the authors proposed a method called ENCORE, which maintains partial trajectories and implements upward exchanges by truncating the trajectories. This method can make MH correction without evaluating intermediate marginal densities. Theoretically, they prove target invariance, uniform ergodicity under bounded-weight assumptions. Their analysis further shows that guidance cancels weight fluctuations. The synthetic, molecular, and image generation experiments support the proposed method.

Critical strengths
  • this paper introduces an interesting method enables replica exchange without evaluating intermediate marginal densities. By replacing upward moves by trajectory truncation, it allows intractable terms to cancel in MH ratio.

  • The paper establishes a clear comparison between SMC and replica exchange with same importance weights. This perspective is interesting and helpful to understand the limitations of existing methods and the benefit of guidance.

Critical weaknesses
  • Some parts of the presentation could be clearer, for example, Figure 1 and 2 are difficult to follow.
Questions for authors
  • Could you justify the assumption in Theorem 3.1(b) that wn−1∣nw_{n-1|n} is uniformly bounded above and away from zero?

  • could you also include the runtime, peak memory and effective number of samples per second? it will be helpful to understand this method.

1 comment
AuthorsOpenReview ↗

Response to Reviewer ssdb

We thank the reviewer for their time and feedback on our manuscript. We address each comment below, providing clarification and additional results where appropriate.

Thm. 3.1(b) Assumption (Q1)

Could you justify the assumption in Theorem 3.1(b) that wn−1∣nw_{n-1|n} is uniformly bounded above and away from zero?

We thank the reviewer for raising this point and apologise for the ambiguity in the main text. Thm. 3.1(b) assumes that the backward weight wn−1∣nw_{n-1|n} is uniformly bounded above and below for each nn; App. C.2.2 makes the strictly positive lower bound explicit and uses it to prove uniform ergodicity. The remark after the proof discusses the scope of this assumption.

At each level nn, these bounds give a path exchange acceptance probability of at least w‾n−1∣n/w‾n−1∣n>0\underline w_{n-1|n}/\overline w_{n-1|n}>0, uniformly over replica states. For the unguided proposal Q=PQ=P, the backward weight reduces to Gn−1/GnG_{n-1}/G_n, so bounded log potentials suffice. For a guided proposal, the backward weight also includes the density ratio dPn−1∣n/dQn−1∣n\mathrm dP_{n-1|n}/\mathrm dQ_{n-1|n}. With bounded log potentials, it suffices for this ratio to be uniformly bounded above and away from zero. The invariance result in part (a) holds under the standing assumptions alone.

Runtime, Memory, and Sampling Efficiency (Q2)

Could you also include the runtime, peak memory and effective number of samples per second?

We measured these for unguided (U) and guided (G) ENCORE, using a single 15-replica ladder on each of the three molecular tempering tasks. Each run used one NVIDIA H100 and retained 49,000 outputs after 1,000 warm-up half-sweeps. Tabs. 1 and 2 below report mean ± sample standard deviation over five seeds per configuration.

Table 1. Runtime and peak GPU tensor memory at 49,000 retained outputs.

TargetRuntime U (h)Runtime G (h)Peak memory U (MiB)Peak memory G (MiB)
ALDP 300 K8.44 ± 0.1013.78 ± 0.0453.7 ± 4.365.8 ± 2.6
A4 500 K8.72 ± 0.1314.45 ± 0.2563.5 ± 7.0109.7 ± 4.4
A6 600 K9.05 ± 0.1814.73 ± 0.1065.3 ± 5.7168.6 ± 5.3

Runtime covers setup through sample storage, including warm-up and guidance.

Table 2. Estimated effective sample size per second for energy and torsion.

TargetEnergy ESS/s UEnergy ESS/s GTorsion ESS/s UTorsion ESS/s G
ALDP 300 K0.0094 ± 0.00290.0974 ± 0.00840.0092 ± 0.00610.0493 ± 0.0091
A4 500 K0.0254 ± 0.00750.0657 ± 0.01130.0199 ± 0.00210.0338 ± 0.0044
A6 600 K0.0393 ± 0.00810.0801 ± 0.01860.0309 ± 0.00990.0417 ± 0.0068

Energy denotes molecular potential energy, and torsion denotes the sine of the first backbone ψ\psi angle. We estimate ESS from autocorrelations of each post-warm-up trace, retaining repeated states. We divide each estimate by that run's end-to-end runtime before averaging across seeds.

Guidance took about 1.6× as long per run and raised mean peak GPU tensor memory from 54–65 MiB to 66–169 MiB. Even so, it increased ESS/s on all three targets, with mean paired gains of 2.07–11.36× for energy and 1.50–7.98× for torsion. Please also see our response to Reviewer 45EX's Q3 for replica exchange diagnostics (restart and round-trip rates) and ESS per denoiser evaluation.


Additional Discussion:

In the original molecular tempering experiments, ENCORE used 15 replica levels and CREPE used 51. At the same output count, ENCORE used about 0.2× as many denoiser evaluations as CREPE (Tab. 3 below). Despite using the fewest evaluations of any method, ENCORE achieved the lowest mean error in most comparisons across three molecules and four metrics (see Tabs. 2 and 10 of our manuscript).

Table 3. Denoiser evaluations per run in the original molecular tempering experiments.

MethodRetained samplesDenoiser evaluations (M)Relative to ENCORE
ENCORE49,0004.551.00×
FKD50,0008.451.86×
FKC50,00010.002.20×
Uncontrolled50,00010.002.20×
RNE50,00020.004.39×
CREPE49,00022.554.95×

Figs. 1 and 2 (W1)

Some parts of the presentation could be clearer, for example, Figure 1 and 2 are difficult to follow.

Thank you for pointing this out! To make Figs. 1 and 2 easier to follow, we added a side-by-side schematic of a single exchange between adjacent levels (Fig. 8 in App. E of the updated manuscript). It shows how CREPE proposes a transition in each direction from the current replica states, whereas ENCORE proposes a path exchange by extending one trajectory and truncating the other.


Thank you again for your review. We would be happy to clarify anything further during the discussion period.

Reviewer_n25tOpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies inference-time control of pretrained diffusion models, including distilled samplers. The goal is to sample from a target obtained by reweighting the model's own output distribution. The paper is motivated by a limitation of existing replica-exchange corrections: they rely on an approximate time reversal, which can introduce persistent sampling bias and may be unavailable for some samplers. ENCORE stores the partial generation trajectory of each replica, so an upward move can be performed by truncating the stored trajectory. The paper establishes target invariance and analyzes how guidance and the choice of replica levels affect sampling efficiency. Experiments on synthetic distributions, molecular sampling, and image generation evaluate accuracy, robustness, and diversity.

Critical strengths
  1. I find the trajectory-based construction conceptually clear and well suited to the stated problem. It avoids the need to simulate an unavailable time reversal while preserving the intended target distribution. The perturbed-sampler and coarse-ladder experiments directly test situations in which the existing correction can become unreliable, providing useful evidence for the motivation behind ENCORE.

  2. The analysis of guided and unguided proposals provides insight beyond the proposed algorithm. It explains why adding more replica levels does not necessarily improve sampling efficiency and how guidance changes this behavior. The connection to sequential Monte Carlo also helps explain how related difficulties appear in different sampling methods.

Critical weaknesses
  1. The molecular-tempering comparisons do not clearly isolate the contribution of the sampling algorithm. ENCORE and FKD use physical-energy reweighting, while several baselines use model-density tempering, and the reported output-processing procedures also differ. These choices may contribute to the observed performance differences. In addition, the uncertainty reported in the main tempering tables comes from subsampling a single run, leaving variation across independent runs unclear. Comparisons using a shared target, consistent output processing, and independent runs would provide stronger evidence for the reported advantages.

  2. The practical benefit of exact replica exchange remains insufficiently demonstrated. SNIS achieves lower reported synthetic errors when evaluated with retained importance weights, while Best-of-64 offers a competitive reward-diversity trade-off without targeting the same distribution. These comparisons leave open when ENCORE is preferable under limited resources. The reported evaluations also mainly use unguided proposals, so the practical gains from guided ENCORE and theory-based replica placement remain unclear. Given the additional trajectory storage and repeated image outputs after thinning, a representative evaluation connecting these design choices to runtime, memory, and effective sampling performance would strengthen the case for the method.

Questions for authors
  1. Could you show whether ENCORE's advantage in molecular sampling persists when methods use a shared target, consistent output processing, and comparable computational budgets? This would help establish how much of the improvement comes from the sampling algorithm itself.

  2. Can a representative experiment show whether guidance and the theory-based choice of replica levels improve ENCORE's practical sampling efficiency? This would connect the theoretical analysis more directly to the method's usefulness in practice.

4 comments
AuthorsOpenReview ↗

Response to Reviewer n25t [1/3]

We thank the reviewer for their time and feedback on our manuscript. We address each comment below, providing clarification and additional results where appropriate.

Molecular Tempering Experiments (Q1, W1)

ENCORE and FKD use physical-energy reweighting, while several baselines use model-density tempering.

We agree with the reviewer and apologise for the confusion. Although Sec. 4.3 distinguishes the two targets, this difference may partly explain the performance gaps in our original comparisons. As suggested, we now compare six samplers under the same physical-energy tilt: unguided ENCORE and FKD, plus guided ENCORE, TDS, and guided and unguided CREPE (Tab. 1 below).

The reported output-processing procedures differ.

We agree that a shared procedure is fairer. We originally followed each method's own output conventions: ENCORE and FKD used endpoint denoising (plus bond-constraint projection on A4/A6), whereas the other baselines returned their final diffusion states. In Tab. 1 below, all methods now use the same output processing.

The uncertainty reported in the main tempering tables comes from subsampling a single run.

We have added four independent runs; Tab. 1 below reports the mean ± standard deviation across all five. We will also run additional seeds for the remaining experiments in the paper.

Comparable computational budgets.

Tab. 1 below now compares methods under the same budget of 4.22M denoiser evaluations per run, including initialisation and warm-up.

Table 1. Molecular sampling through physical-energy tilting. U/G denotes unguided/guided variants.

SystemMethodEnergy TVD ↓Distance TVD ↓Aligned W2\mathrm{W}_2 ↓Torsion MMD2\rm MMD^2 ↓
ALDPENCORE-G0.1689 ± 0.00970.0066±0.0003‾\underline{0.0066±0.0003}0.1717±0.0023‾\underline{0.1717±0.0023}0.00075 ± 0.00026
ALDPENCORE-U0.1955 ± 0.03420.0145 ± 0.00150.2373 ± 0.00400.00527 ± 0.00325
ALDPCREPE-G0.2356 ± 0.01400.0095 ± 0.00070.1932 ± 0.00360.00471 ± 0.00087
ALDPCREPE-U0.1842 ± 0.01650.0099 ± 0.00080.2173 ± 0.00210.00292 ± 0.00108
ALDPTDS0.1439±0.0064‾\underline{0.1439±0.0064}0.0063 ± 0.00040.1635 ± 0.00050.00101±0.00041‾\underline{0.00101±0.00041}
ALDPFKD0.1260 ± 0.00720.0107 ± 0.00060.2017 ± 0.00210.00508 ± 0.00147
A4ENCORE-G0.1418 ± 0.00750.0096 ± 0.00110.6955±0.0030‾\underline{0.6955±0.0030}0.00221 ± 0.00036
A4ENCORE-U0.1502 ± 0.01720.0102 ± 0.00050.7348 ± 0.00260.00289 ± 0.00087
A4CREPE-G0.3159 ± 0.00700.0129 ± 0.00120.7291 ± 0.00350.00704 ± 0.00177
A4CREPE-U0.2411 ± 0.00960.0104 ± 0.00070.7295 ± 0.00380.00304 ± 0.00070
A4TDS0.0953±0.0021‾\underline{0.0953±0.0021}0.0082 ± 0.00080.6814 ± 0.00160.00112 ± 0.00018
A4FKD0.0885 ± 0.01100.0088±0.0009‾\underline{0.0088±0.0009}0.7137 ± 0.00190.00145±0.00021‾\underline{0.00145±0.00021}
A6ENCORE-G0.1067 ± 0.00480.0078±0.0007‾\underline{0.0078±0.0007}1.2065±0.0028‾\underline{1.2065±0.0028}0.00135 ± 0.00015
A6ENCORE-U0.1107 ± 0.01230.0089 ± 0.00131.2614 ± 0.00860.00186 ± 0.00023
A6CREPE-G0.3055 ± 0.01060.0079 ± 0.00061.2530 ± 0.00190.00271 ± 0.00048
A6CREPE-U0.2256 ± 0.01080.0076 ± 0.00081.2511 ± 0.00550.00185 ± 0.00019
A6TDS0.0682 ± 0.00530.0081 ± 0.00041.1880 ± 0.00180.00103 ± 0.00008
A6FKD0.0717±0.0020‾\underline{0.0717±0.0020}0.0084 ± 0.00071.2352 ± 0.00210.00118±0.00012‾\underline{0.00118±0.00012}

Guided ENCORE (Q2, W2)

Can a representative experiment show whether guidance and the theory-based choice of replica levels improve ENCORE's practical sampling efficiency?

In Tab. 1 above, guided ENCORE has lower mean error than unguided ENCORE on all four metrics for all three molecules. Please also see our response on Runtime, Memory, and Sampling Efficiency (W2) below.

AuthorsOpenReview ↗

Response to Reviewer n25t [2/3]

Exactness (W2)

The practical benefit of exact replica exchange remains insufficiently demonstrated.

Exactness is our main motivation, and the perturbed GMM experiment (Sec. 4.1, Fig. 4) already illustrates its benefit, where ENCORE's W2\mathrm{W}_2 error is 0.0150.015, versus 0.6500.650 for CREPE.

We also added a controlled synthetic experiment with an OU generator and an analytically known Gaussian-mixture target. Holding both fixed, we vary the assumed reverse-proposal variance. Tab. 2 below reports a region probability (true value 0.50.5) after 64 exchange cycles, as mean ± standard deviation over five seeds. As the mismatch grows, CREPE's estimate drops to 0.300.30, while ENCORE and CREPE + exact MH (an oracle reference, intractable for learned models) stay near 0.50.5.

Table 2. Target preservation under controlled reversal mismatch.

Reverse-proposal variance multiplierCREPECREPE + Exact MH*ENCORE
1.00 (matched reversal)0.4995 ± 0.00160.4995 ± 0.00160.5000 ± 0.0063
0.750.3961 ± 0.00490.4986 ± 0.00420.5000 ± 0.0063
0.500.2971 ± 0.00480.4978 ± 0.00690.5000 ± 0.0063

*CREPE + exact MH keeps CREPE's proposals but uses the known target and proposal densities to isolate the acceptance correction. This oracle is generally intractable for learned diffusion models, whose intermediate target densities are unavailable.

SNIS and BoN (W2)

SNIS achieves lower reported synthetic errors when evaluated with retained importance weights.

We agree and should have stated this more clearly. SNIS works well when the proposal covers the target, but coverage can deteriorate under stronger tilts or in higher dimensions.

To test this in practice, we compared SNIS and ENCORE on eight tasks: a 1D Gaussian shifted by 8σ8\sigma, three peptide tempering tasks, and four BioEmu protein benchmarks biased towards lower native-contact fractions (see our response to Reviewer 45EX, Q2). Guided ENCORE had lower mean error than SNIS on all eight (Tab. 3 below).

Table 3. Sampling accuracy of SNIS and ENCORE.

Experiment (Metric ↓)SNISUnguided ENCOREGuided ENCORE
Gaussian shift (CDF Error)0.9533 ± 0.02020.8729±0.0474‾\underline{0.8729±0.0474}0.0248 ± 0.0108
ALDP 300 K (Energy TVD)0.6289 ± 0.03300.1955±0.0342‾\underline{0.1955±0.0342}0.1689 ± 0.0097
A4 500 K (Energy TVD)0.2715 ± 0.02280.1502±0.0172‾\underline{0.1502±0.0172}0.1418 ± 0.0075
A6 600 K (Energy TVD)0.1793 ± 0.01570.1107±0.0123‾\underline{0.1107±0.0123}0.1067 ± 0.0048
1ENH (RMSD W1\rm W_1)62.14 ± 33.6342.59±6.36‾\underline{42.59±6.36}37.33 ± 12.99
1BTA (RMSD W1\rm W_1)2.32 ± 1.271.98±1.96‾\underline{1.98±1.96}1.36 ± 0.33
1TTG (RMSD W1\rm W_1)63.55 ± 103.0911.29±9.62‾\underline{11.29±9.62}5.48 ± 1.70
2RN2 (RMSD W1\rm W_1)8.24 ± 2.346.95 ± 0.878.20±1.46‾\underline{8.20±1.46}

Entries are mean ± standard deviation over 32 seeds (Gaussian shift), five (molecular tempering) or three (proteins).

Best-of-64 offers a competitive reward-diversity trade-off without targeting the same distribution.

We agree that Best-of-N (BoN) offers a strong reward–diversity trade-off on the image tasks. For each prompt, BoN returns the highest-reward image from each of 32 independent groups of 64 draws, so its outputs are independent and distinct. ENCORE instead records two outputs from each of 16 independent chains; the two coincide when the chain's target-level replica has not changed, lowering uniqueness and diversity scores. At a fixed budget, fewer but longer chains trade diversity for reward: with 8 chains instead of 16, mean ImageReward rises from 1.711 to 1.745, while DINO-Vendi falls from 3.02 to 2.78 (Tab. 13 of the updated manuscript).

Unlike BoN, ENCORE and the Monte Carlo baselines target q0∝exp⁡[r] p0q_0\propto\exp[r]\ p_0. However, these metrics cannot assess target fidelity for any method: reward and generic diversity do not measure discrepancy from q0q_0, and no reliable target reference is available. We therefore interpret the image results as showing that ENCORE is competitive and works with distilled samplers, including DMD2, whose native sampler lacks the noising kernel CREPE requires. We will clarify this scope in the manuscript.

AuthorsOpenReview ↗

Response to Reviewer n25t [3/3]

Runtime, Memory, and Sampling Efficiency (W2)

Given the additional trajectory storage ... a representative evaluation connecting these design choices to runtime, memory, and effective sampling performance would strengthen the case for the method.

We agree that ENCORE incurs an additional storage cost, as noted in our limitations (Sec. 5). For NN replica levels and dd-dimensional states, storing the paths requires O(N2d)O(N^2d) sampler-state memory, compared with O(Nd)O(Nd) for CREPE and O(Md)O(Md) for SMC methods such as TDS with MM parallel particles. In exchange, the stored paths let ENCORE move up the ladder by truncation, without additional denoiser evaluations.

ENCORE can be preferable when memory or parallelism is fixed and samples are needed online. Its path storage stays constant as samples accumulate. Counting states at replica levels, a single ENCORE ladder stores fewer states than MM parallel SMC particles when M>N2M>N^2. In our molecular tempering experiments, N=15N=15 and each run retains M=49,000M=49{,}000 samples, so N2=225≪MN^2=225\ll M.

In the original molecular tempering experiments, ENCORE used 15 replica levels and CREPE used 51. At the same output count, ENCORE used about 0.2× as many denoiser evaluations as CREPE (Tab. 4 below). Despite using the fewest evaluations of any method, ENCORE achieved the lowest mean error in most comparisons across three molecules and four metrics (see Tabs. 2 and 10 of our manuscript).

Table 4. Denoiser evaluations per run in the original molecular tempering experiments.

MethodRetained samplesDenoiser evaluations (M)Relative to ENCORE
ENCORE49,0004.551.00×
FKD50,0008.451.86×
FKC50,00010.002.20×
Uncontrolled50,00010.002.20×
RNE50,00020.004.39×
CREPE49,00022.554.95×

To assess practical sampling efficiency, we compare unguided (U) and guided (G) ENCORE on the same ladder. Each run used one NVIDIA H100 and retained 49,000 outputs after 1,000 warm-up half-sweeps. Tab. 5 below reports mean ± sample standard deviation over five seeds per configuration.

Table 5. Runtime, peak GPU memory and estimated energy ESS at 49,000 outputs.

TargetRuntime U (h)Runtime G (h)Peak memory U (MiB)Peak memory G (MiB)Energy ESS UEnergy ESS G
ALDP 300 K8.44 ± 0.1013.78 ± 0.0453.7 ± 4.365.8 ± 2.6285 ± 884,830 ± 410
A4 500 K8.72 ± 0.1314.45 ± 0.2563.5 ± 7.0109.7 ± 4.4797 ± 2343,419 ± 600
A6 600 K9.05 ± 0.1814.73 ± 0.1065.3 ± 5.7168.6 ± 5.31,283 ± 2794,252 ± 1,012

Runtime covers setup through sample storage, including warm-up and guidance. We estimate ESS from autocorrelations of each post-warm-up trace, retaining repeated states.

Guidance took about 1.6× as long per run, yet improved energy ESS per second by 11.36× on ALDP, 2.80× on A4 and 2.07× on A6 (mean within-seed gain). Mean gains in torsion ESS per second, measured using the sine of the first backbone ψ\psi angle, ranged from 1.50× on A6 to 7.98× on ALDP. We have updated our manuscript to report absolute ESS per second and its variability across seeds (App. F.5.3, Tab. 17).


Thank you again for your review. We would be happy to clarify anything further during the discussion period.

Thank you for the additional experiments. The authors have substantially addressed my concerns throughout the rebuttal, and I will raise my rating accordingly.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to All Reviewers

We sincerely thank the reviewers and the Area Chair for their time and consideration. We appreciate the reviewers' constructive feedback and have revised the manuscript accordingly; changes are highlighted in blue, and the new results are in App. F.3.1 and F.5.

To address the concerns, we added the following experiments:

  • A controlled test of exactness with a known target (n25t, Tab. 2)
  • A matched comparison of six samplers, including guided ENCORE (n25t, Tab. 1)
  • Comparisons with SNIS on eight tasks (n25t, Tab. 3)
  • Four BioEmu proteins with up to 930 dimensions (45EX, Tabs. 2–3)
  • Runtime, memory and sampling-efficiency diagnostics (ssdb, Tabs. 1–2; 45EX, Tabs. 4–5)

Guided ENCORE is better than both CREPE variants in 39 of 40 peptide and protein comparisons and, unlike CREPE, preserves the target exactly (n25t, Tab. 2). SMC methods such as TDS remain strong at matched budgets; ENCORE provides their exact replica-exchange counterpart.

For memory, ENCORE stores O(N2d)O(N^2d) states, independent of the number of samples. With guided proposals, about N≈2ΛN\approx 2\Lambda levels suffice for stable communication (Thm. 3.2), whereas CREPE needs a finer ladder because the error in its acceptance ratio builds up over the model steps that an exchange covers.

Thank you all again. We look forward to the discussion and would be happy to clarify any remaining points.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.