acceptodds
Under review as a conference paper at ICLR 2027

Error-Corrected Inference-Time Scaling for Imperfect Diffusion Models

10 views

Abstract

Inference-time scaling adapts pretrained diffusion models to new sampling tasks without additional training. Existing methods rely primarily on Monte Carlo sampling with more particles, yet are premised on the pretrained model being exact. In practice, data and training limitations make the model imperfect, and these methods inherit its error. More particles reduce Monte Carlo error but cannot remove the mismatch between the endpoint and the desired target or the error in tracking the prescribed probability path. We introduce the Energy-based Feynman-Kac Corrector (EBFKC), a framework for energy-based diffusion models that corrects these errors on the fly given a reference energy. We first derive Feynman-Kac dynamics that track a prescribed path exactly in the continuous-time population limit even when the model is imperfect, and approximate these dynamics using sequential Monte Carlo with variance-controlling guidance. To remove the endpoint mismatch, we use the pretrained energy as a surrogate along the diffusion path and progressively incorporate the discrepancy between the learned and target terminal energies. Experiments on Gaussian mixture models, particle systems, alanine dipeptide, and alanine tetrapeptide show that our method closely matches target distributions and molecular free-energy profiles under annealing and reward tilting, whereas standard inference-time scaling baselines retain substantial sampling errors.

Primary area:
generative models
Keywords:
Diffusion Models, Inference-time Scaling, Energy-based Modeling

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_idu7OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies inference-time sampling when a pretrained diffusion model is imperfect and an external reference energy specifies the desired target distribution. It distinguishes two sources of error: endpoint mismatch, where the learned terminal density differs from the reference target, and path-tracking error, where sampling dynamics fail to follow the prescribed intermediate densities. The authors introduce the Energy-based Feynman–Kac Corrector (EBFKC), a framework for energy-based diffusion models that corrects these errors on the fly given a reference energy. Experiments cover Gaussian mixtures, interacting particle systems, and molecular sampling and free-energy estimation.

Critical strengths
  1. The paper correctly distinguishes the density defined by a learned energy from the marginal distribution generated by the corresponding reverse SDE. These need not coincide. This distinction explains why increasing the particle count alone cannot repair every source of sampling error, and provides a useful framework for interpreting inference-time correction methods. 2.The reported improvements are substantial across several relevant benchmarks.The experiments extend beyond analytical examples to learned molecular energy models. For ALDP annealing, Table 2 reports a reduction in Ramachandran TVD from 0.50630.5063 for EBFKC on MIT to 0.08550.0855 for EBFKC on EAT, and a reduction in basin-population error from 0.16570.1657 to 0.00470.0047. The free-energy results in Figure 5 are also encouraging.
Critical weaknesses

The experiments diagnose the proposed failure modes more convincingly than they establish a practical advantage over competing correction strategies. The geometric tempering baseline on LJ-13 is useful, but its comparison uses the same particle count rather than a demonstrated comparable computational budget. Its runtime is also absent from Table 8. Across the paper, runtime tables report individual operating points rather than accuracy as a function of computational budget, despite the inference-time scaling motivation. The derivative overhead is material: Table 9 reports approximately 508508 seconds for EBFKC on EAT versus 143143 seconds for EBFKC on MIT.

Questions for authors

Given the unconstrained architecture with sampling-time endpoint replacement in Grenioux and Noble, what precise capability or guarantee distinguishes EBFKC beyond the progressive correction schedule and its integration with VCG?

3 comments
AuthorsOpenReview ↗

Response to Reviewer idu7 [1/3]

We sincerely thank you for highlighting the strengths and novelty of our work, and for your insightful suggestions! We are glad that you find that our paper "correctly distinguishes the density defined by a learned energy from the marginal distribution generated by the corresponding reverse SDE." As you note, this distinction "explains why increasing the particle count alone cannot repair every source of sampling error, and provides a useful framework for interpreting inference-time correction methods." We also appreciate your observation that "the reported improvements are substantial across several relevant benchmarks" and that "the experiments extend beyond analytical examples to learned molecular energy models." Thank you as well for noting that "the free-energy results in Figure 5 are also encouraging." We address your suggestions below.

Practical advantage over competing strategies under matched computational budgets

"The experiments diagnose the proposed failure modes more convincingly than they establish a practical advantage over competing correction strategies."

We find this suggestion really helpful. We have added matched-budget comparisons to clarify that EBFKC has a practical advantage over competing strategies.

(a) Tempering SMC on LJ-13

"The geometric tempering baseline on LJ-13 is useful, but its comparison uses the same particle count rather than a demonstrated comparable computational budget. Its runtime is also absent from Table 8."

For tempering-path SMC on LJ-13, we scale the number of particles and report wall-clock times:

  • Annealing (γ\gamma=2.0)
MethodParticlesWall Clock (s)∣ΔNLL∣\lvert\Delta\mathrm{NLL}\rvert ↓MMD ↓SWD ↓RDF W1 ↓Energy W1 ↓
Tempering SMC10,0002.93 ± 0.0517.2893 ± 1.87200.4850 ± 0.10480.4328 ± 0.05840.2328 ± 0.020217.29 ± 1.87
100,00012.45 ± 0.0217.1848 ± 2.31450.2296 ± 0.08480.2214 ± 0.03000.2195 ± 0.040117.18 ± 2.31
200,00021.98 ± 0.0416.2489 ± 0.98580.1820 ± 0.01000.2056 ± 0.03810.2058 ± 0.012716.25 ± 0.99
500,00050.90 ± 0.0915.0536 ± 1.20250.2054 ± 0.07440.2035 ± 0.06810.1974 ± 0.018515.05 ± 1.20
1,000,00099.09 ± 0.0615.5165 ± 1.53870.1296 ± 0.03000.1243 ± 0.02810.2058 ± 0.016915.52 ± 1.54
EBFKC on EAT (Ours)8,1921339.71 ± 0.450.9211 ± 0.39940.0600 ± 0.00590.0647 ± 0.00980.0121 ± 0.00750.94 ± 0.40
  • Reward tilting (λ′\lambda'=0.5)
MethodParticlesWall Clock (s)∣ΔNLL∣\lvert\Delta\mathrm{NLL}\rvert ↓MMD ↓SWD ↓RDF W1 ↓Energy W1 ↓
Tempering SMC10,0002.97 ± 0.0613.7816 ± 3.43720.3175 ± 0.09500.2809 ± 0.09720.3886 ± 0.073913.78 ± 3.44
100,00012.92 ± 0.0314.7314 ± 1.01560.1656 ± 0.02200.1749 ± 0.03630.4109 ± 0.030914.73 ± 1.02
200,00022.87 ± 0.0513.9917 ± 0.74620.1465 ± 0.01360.1541 ± 0.02210.3805 ± 0.032113.99 ± 0.75
500,00052.93 ± 0.0914.4294 ± 0.61130.1297 ± 0.00210.1347 ± 0.00940.4050 ± 0.009314.43 ± 0.61
1,000,000103.21 ± 0.0714.2269 ± 0.59240.1201 ± 0.00310.1327 ± 0.00990.3904 ± 0.029214.23 ± 0.59
EBFKC on EAT (Ours)8,1921820.83 ± 1.570.1366 ± 0.08320.0345 ± 0.00410.0425 ± 0.00230.0140 ± 0.00240.25 ± 0.08

Tempering SMC is much faster per particle, but its error barely changes over a 100-fold increase in particles (∣ΔNLL∣|\Delta\mathrm{NLL}| from 17.3 to 15.5 for annealing and from 13.8 to 14.2 for reward tilting) and remains more than an order of magnitude above that of EBFKC on EAT even at 1,000,000 particles.

AuthorsOpenReview ↗

Response to Reviewer idu7 [2/3]

(b) Accuracy as a function of compute budget, and derivative overhead (ALDP)

"Across the paper, runtime tables report individual operating points rather than accuracy as a function of computational budget, despite the inference-time scaling motivation. The derivative overhead is material: Table 9 reports approximately (508) seconds for EBFKC on EAT versus (143) seconds for EBFKC on MIT."

For ALDP annealing, we compare particle-count scaling for EBFKC on EAT and EBFKC on MIT. To approximately match compute budgets, EBFKC on MIT uses four times as many particles per batch as EBFKC on EAT. Results are evaluated on samples accumulated over 25 batches:

MethodBatch-wise particlesWall Clock (s)Energy TVD ↓Distance TVD ↓
EBFKC on EAT100100.95 ± 2.450.0771800.009166
200140.68 ± 2.240.0472070.007234
500308.38 ± 2.500.0351130.004006
1,000613.05 ± 5.040.0235550.002973
EBFKC on MIT40076.38 ± 0.600.4421420.038203
800147.05 ± 0.820.4378520.035386
2,000349.93 ± 1.570.4355230.037053
4,000695.40 ± 2.330.4559240.037037

We have also included the inference-time scaling curve in Fig. 5 of the revised paper. Scaling the number of particles clearly improves EBFKC on EAT. EBFKC on MIT does not benefit from more particles because of the endpoint mismatch error. The overhead you note in Table 9 of the paper (508 s vs. 143 s) comes mainly from the OpenMM finite differences for the reference-energy terms, which are not parallelized across particles (App. B.4; see also our complexity analysis in the response to Reviewer aQ3B); even so, at a matched wall-clock time EBFKC on EAT is far more accurate than EBFKC on MIT with more particles.

AuthorsOpenReview ↗

Response to Reviewer idu7 [3/3]

Comparison with DABG

"Given the unconstrained architecture with sampling-time endpoint replacement in Grenioux and Noble, what precise capability or guarantee distinguishes EBFKC beyond the progressive correction schedule and its integration with VCG?"

We thank the reviewer for this insightful question, which lets us clarify how EBFKC differs from DABG [1]. The main differences are:

  • Setting and generality. We target inference-time scaling, i.e., sampling qTref∝(pref)γerq_T^{\mathrm{ref}} \propto (p^{\mathrm{ref}})^\gamma e^{r} under arbitrary annealing and reward tilting, rather than the equilibrium distribution p⋆p^\star alone. DABG is therefore a particular case of EBFKC:
    • its pinned architecture corresponds to EBFKC on the model-induced path with γ=1\gamma=1 and r=0r=0 (Sec. 3.3);
    • its unconstrained architecture with sampling-time endpoint replacement corresponds to an EAT path whose correction schedule κt\kappa_t jumps from 0 to 1 at t=Tt=T.
  • General prescribed paths, purely post-training. Theorem 2 guarantees exact tracking of any prescribed path in the continuous-time population limit, not only the raw energy-based marginals. We exploit this freedom with a look-ahead map that avoids evaluating noisy samples directly on ErefE^{\mathrm{ref}} and Eθ(⋅,T)E_\theta(\cdot,T); without it, our ALDP runs become numerically unstable (App. D.4). DABG instead uses either the raw energy path or a pinned architecture, which can be seen as a look-ahead built into training. Our construction requires no change to training.
  • Progressive correction schedule and VCG. As you note, we additionally integrate a progressive correction schedule with VCG, and both proved essential in our experiments. On the hardest molecular problems, an appropriate schedule is needed for accurate sampling, and VCG is key to weight stability when the pretrained model is imperfect. Neither is explored in DABG.
  • Complementary tools in DABG. DABG also explores a replica-exchange-based correction, which does not directly transfer to our framework. It also uses integration schemes beyond Euler–Maruyama, which could readily be combined with EBFKC.

We thank you again for your constructive feedback. We hope our responses and additional experimental results have addressed your concerns, and we would be very grateful if you could consider raising your score.


Reference

[1] Grenioux, Louis, and Maxence Noble. "Diffusion-based Annealed Boltzmann Generators: benefits, pitfalls and hopes." arXiv preprint arXiv:2601.21026 (2026).

Reviewer_aQ3BOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes the Energy-based Feynman–Kac Corrector (EBFKC), a framework for correcting the effects of model training errors on inference-time scaling in energy-based diffusion models. It assumes access to the reference distribution's energy and uses this information during inference to correct an imperfect pretrained model. The learned energy model guides sampling, while the reference energy is incorporated into the prescribed path and reweighting to correct model error. The framework also incorporates variance-controlling guidance. Experiments demonstrate improved sampling accuracy for annealed and reward-tilted targets.

Critical strengths

The paper addresses a meaningful limitation of inference-time scaling derived under the assumption of a perfect model. In scientific sampling problems where a reference energy is available, using this information to correct the learned imperfect model is useful.

The controlled GMM experiments are particularly informative. They illustrate the contribution of each component—EBFKC and EAT—across different levels of model error and demonstrate the effectiveness of the proposed method even when the model error is large.

Critical weaknesses

Gap between the paper's framing and its actual scope and contribution. The title does not reflect the restriction to energy-based diffusion models or the need for access to the reference energy, making it broader than the proposed framework. EBFKC would also be better framed as removing FKC's simplifying assumption, rather than as a method derived from a new Feynman–Kac-type Fokker–Planck equation. Under suitable regularity and integrability conditions, FK formulations that combine prescribed marginal paths with different proposal dynamics are already well established in the SMC literature. A key contribution of FKC (Skreta et al., 2025) is to simplify the computation of the reweighting potential by assuming that the pretrained model's marginal densities satisfy the base Fokker–Planck equation. EBFKC essentially removes this assumption and evaluates the full reweighting potential to correct for violations of it, at the cost of additional derivative computations.

The scalability of the method is insufficiently established. The method is evaluated on relatively small systems. Although the results on GMM, DW-4, LJ-13, ALDP, and alanine tetrapeptide are interesting, they do not sufficiently establish its practical significance for larger systems. Evaluation on larger systems studied in recent work on Boltzmann generators would strengthen the evidence for its practical relevance. In addition, estimating neural-network Laplacians introduces potentially substantial computational costs. The reported experiments do not establish whether these costs remain manageable as the system size increases.

Questions for authors

Q1. Could you evaluate the method on larger systems, such as hexa-alanine or chignolin?

Q2. Could you clarify the computational complexity of EBFKC relative to FKC? In particular, EBFKC requires additional derivative evaluations, including the Laplacian of the learned energy. Although empirical runtimes are reported, a complexity comparison accounting for the Laplacian would help clarify the method's computational limitations and scaling behavior.

Q3. Could you include FKC on MIT in the quantitative GMM and particle-system comparisons to provide a direct comparison with the existing method?

Q4. Could you compare EBFKC with FKC using more particles under matched compute budgets? This would clarify whether EBFKC offers better compute-normalized sampling performance.

7 comments
AuthorsOpenReview ↗

Response to Reviewer aQ3B [1/3]

We sincerely thank you for your insightful feedback and detailed review, and for recognizing our contribution to addressing the limitations of existing inference-time scaling methods. We are glad that you find that our paper "addresses a meaningful limitation of inference-time scaling derived under the assumption of a perfect model." We are also glad you agree that "in scientific sampling problems where a reference energy is available, using this information to correct the learned imperfect model is useful." Finally, we appreciate your view that "the controlled GMM experiments are particularly informative," illustrating "the contribution of each component—EBFKC and EAT—across different levels of model error." We address your suggestions below.

Gap between the paper's framing and its actual scope and contribution

"The title does not reflect the restriction to energy-based diffusion models or the need for access to the reference energy, making it broader than the proposed framework. EBFKC would also be better framed as removing FKC's simplifying assumption, rather than as a method derived from a new Feynman–Kac-type Fokker–Planck equation. […] EBFKC essentially removes this assumption and evaluates the full reweighting potential to correct for violations of it, at the cost of additional derivative computations."

We agree with your suggestion to frame EBFKC as removing FKC's simplifying assumption; this is precisely why we call our method "energy-based FKC." Following your suggestion, we have changed the title to "Error-Corrected Inference-Time Scaling for Imperfect Energy-Based Diffusion Models" in the revised paper, which better reflects the scope of our framework.

Regarding access to the reference energy, it is needed only to correct the endpoint mismatch, not to run EBFKC itself. Without an external reference energy, e.g., when the goal is to steer the model's own distribution, one sets Eref=Eθ(⋅,T)E^{\mathrm{ref}}=E_\theta(\cdot,T), so the model-induced path already ends at the desired target and there is no endpoint mismatch (Sec. 2.2). EBFKC on this path (EBFKC on MIT) still removes the path-tracking error of FKC, because its weights require only the learned energy (Theorem 2). Our experiments isolate this effect with FKC on EAT, which redefines the reward so that its target coincides with qTrefq_T^{\mathrm{ref}} and thus removes the endpoint mismatch while retaining FKC's path-tracking error. It remains substantially less accurate than EBFKC on EAT (e.g., GMM annealing ∣ΔNLL∣|\Delta\mathrm{NLL}| 1.0431.043 vs. 0.1330.133), and the gap grows with the model error in our GMM error sweep. The new title therefore names the energy-based restriction, and the error correction it refers to applies with or without a reference energy, while the abstract and introduction state that removing the endpoint mismatch requires one.

We also agree that Feynman–Kac formulations combining prescribed marginal paths with different proposal dynamics are well established in the SMC literature. Within this family, the key contribution of FKC [1] is to build on the diffusion model's own dynamics and to make the weights tractable through the Fokker–Planck equation of the base diffusion, so that they depend only on the score. Our approach instead takes the marginal densities of an energy-based diffusion model as the prescribed path. The proposal is the reverse-diffusion drift that would generate this path exactly if it satisfied the ideal Fokker–Planck equation. We make two remarks:

  1. Tractable weights without a consistency assumption. With this energy-based formulation, the full reweighting potential is tractable without any consistency assumption. FKC's weights, by contrast, are tractable only under the assumption that the model satisfies the base Fokker–Planck equation.
  2. A different proposal from the standard route. There are many ways to build FK dynamics for a given path of densities. A popular one is annealed Langevin dynamics, which we detail in App. B.7 and use as our tempering SMC baseline. Our diffusion-based proposal departs from this standard route, and it substantially outperforms tempering SMC on LJ-13 (Sec. 4.2).
AuthorsOpenReview ↗

Response to Reviewer aQ3B [2/3]

Computational complexity of EBFKC

"In addition, estimating neural-network Laplacians introduces potentially substantial computational costs. The reported experiments do not establish whether these costs remain manageable as the system size increases. [...] Could you clarify the computational complexity of EBFKC relative to FKC? In particular, EBFKC requires additional derivative evaluations, including the Laplacian of the learned energy. Although empirical runtimes are reported, a complexity comparison accounting for the Laplacian would help clarify the method's computational limitations and scaling behavior.

We thank the reviewer for the insightful suggestion to analyze the computational complexity, which is directly related to the method's scalability.

Let NN and KK denote the number of particles and sampling steps, and let CC be the cost of one gradient evaluation of the energies, which depends on the network or force field rather than on our method. A Hessian-vector product costs a small constant multiple of a gradient, and the finite-difference Hutchinson estimate of the Laplacian (App. B.4) needs two gradients per probe. Hence, on the same path, FKC costs O(NKC)O(NKC) and EBFKC costs O(NKRC)O(NKRC), where RR is the number of Hutchinson probes. The per-gradient cost CC grows with the dimension dd (e.g., quadratically for pairwise architectures), but it enters FKC and EBFKC alike, so the relative overhead of EBFKC depends on dd mainly through RR, which must grow with dimension to keep the variance of the Laplacian estimate under control (R=2R=2 for DW-4 up to R=8R=8 for Ala-4 in our experiments). Hutch++ ([2]; App. B.4) reduces the number of probes needed for a given accuracy and can mitigate this growth.

Empirically (runtime tables in App. D), the extra Laplacian of EBFKC is cheap: on the same path, EBFKC and FKC have similar runtime (within about 15% on GMM, DW-4, and LJ-13). Most of the overhead instead comes from evaluating the reference energy along the EAT path, which FKC on EAT, the baseline that also uses the reference energy along the path, pays as well; in our experiments EBFKC on EAT takes 1.1×1.1\times to 3.6×3.6\times the runtime of EBFKC on MIT, and 1.3×1.3\times on Ala-4, our largest system.

Scalability to larger systems

"The scalability of the method is insufficiently established. The method is evaluated on relatively small systems. […] Evaluation on larger systems studied in recent work on Boltzmann generators would strengthen the evidence for its practical relevance. [...] Could you evaluate the method on larger systems, such as hexa-alanine or chignolin?"

Across all systems we consider, from GMMs to the 42-atom Ala-4, EBFKC on EAT shows a clear advantage over existing methods. This advantage is structural: the errors of FKC- and MIT-based baselines do not vanish with more particles (see our bound in the response to S8wq and the matched-compute results), whereas EBFKC on EAT targets the reference distribution by construction.

On larger systems, pretrained models can be less accurate, and our GMM sweep (Fig. 2c,d) shows the gap widening as model error grows, so we expect the correction to matter more there.

We are also optimistic about the remaining practical hurdles. Machine-learned interatomic potentials can replace the OpenMM finite differences (App. B.4), Hutch++ can improve the Laplacian estimates, and correction schedules can be adapted automatically (Sec. 5). This makes systems such as hexa-alanine and chignolin a promising next step. We have started setting up experiments on hexa-alanine and will share the results in this thread if they become available during the discussion period.

AuthorsOpenReview ↗

Response to Reviewer aQ3B [3/3]

Quantitative comparisons with FKC on MIT

"Could you include FKC on MIT in the quantitative GMM and particle-system comparisons to provide a direct comparison with the existing method?"

We have added FKC on MIT to the quantitative comparisons on GMM and the particle systems. Its results are compared directly with EBFKC on EAT below.

GMM

  • Annealing (γ\gamma=2.0)
Method∣ΔNLL∣\lvert\Delta\mathrm{NLL}\rvert ↓MMD ↓SWD ↓Mean L2 ↓Cov F / 10310^3 ↓
FKC on MIT59.646 ± 1.7060.091 ± 0.0012.978 ± 0.5089.056 ± 1.0441.715 ± 0.080
EBFKC on EAT (Ours)0.133 ± 0.0950.022 ± 0.0011.110 ± 0.3315.493 ± 0.7740.639 ± 0.028
  • Reward tilting (σ\sigma=100)
Method∣ΔNLL∣\lvert\Delta\mathrm{NLL}\rvert ↓MMD ↓SWD ↓Mean L2 ↓Cov F / 10310^3 ↓
FKC on MIT23.036 ± 6.3150.374 ± 0.04710.126 ± 5.62661.931 ± 24.9311.028 ± 1.385
EBFKC on EAT (Ours)0.113 ± 0.0160.018 ± 0.0080.194 ± 0.1021.000 ± 0.6840.050 ± 0.033

DW-4

  • Annealing (γ\gamma=2.0)
Method∣ΔNLL∣\lvert\Delta\mathrm{NLL}\rvert ↓MMD ↓SWD ↓RDF W1 ↓Energy W1 ↓
FKC on MIT70.5989 ± 0.53820.1633 ± 0.00180.3762 ± 0.01991.0848 ± 0.004870.5989 ± 0.5382
EBFKC on EAT (Ours)0.3389 ± 0.21070.0816 ± 0.01930.2032 ± 0.04410.3943 ± 0.14370.3754 ± 0.1556
  • Reward tilting (λ′\lambda'=0.5)
Method∣ΔNLL∣\lvert\Delta\mathrm{NLL}\rvert ↓MMD ↓SWD ↓RDF W1 ↓Energy W1 ↓
FKC on MIT53.8794 ± 0.43800.1538 ± 0.00180.3730 ± 0.02641.0833 ± 0.005453.8794 ± 0.4380
EBFKC on EAT (Ours)0.3769 ± 0.05680.0178 ± 0.00280.0461 ± 0.00520.0690 ± 0.00650.3769 ± 0.0568

LJ-13

  • Annealing (γ\gamma=2.0)
Method∣ΔNLL∣\lvert\Delta\mathrm{NLL}\rvert ↓MMD* ↓SWD* ↓RDF W1 ↓Energy W1 ↓
FKC on MIT14.2786 ± 0.14640.0200 ± 0.00090.0212 ± 0.00250.0424 ± 0.000914.2786 ± 0.1464
EBFKC on EAT (Ours)0.9211 ± 0.39940.0600 ± 0.00590.0647 ± 0.00980.0121 ± 0.00750.9411 ± 0.3991
  • Reward tilting (λ′\lambda'=0.5)
Method∣ΔNLL∣\lvert\Delta\mathrm{NLL}\rvert ↓MMD* ↓SWD* ↓RDF W1 ↓Energy W1 ↓
FKC on MIT21.1555 ± 5.58150.0171 ± 0.00100.0206 ± 0.00250.0274 ± 0.000821.1555 ± 5.5815
EBFKC on EAT (Ours)0.1366 ± 0.08320.0345 ± 0.00410.0425 ± 0.00230.0140 ± 0.00240.2454 ± 0.0837

*: This discrepancy arises because MMD and SWD may not fully capture local structural errors, such as particle overlaps or steric clashes, that can lead to substantial deviations in energy.

These results show that FKC on MIT suffers from path-tracking error and endpoint mismatch at the same time, which leads to substantial sampling error.


Comparing EBFKC with FKC under matched compute budgets

"Could you compare EBFKC with FKC using more particles under matched compute budgets? This would clarify whether EBFKC offers better compute-normalized sampling performance."

We agree that comparing EBFKC and FKC under matched computational budgets gives a fairer assessment of their relative sampling efficiency. Taking LJ-13 annealing as an example, we increase the number of particles for FKC on MIT to match the compute budget of EBFKC on EAT. The results are below:

MethodParticlesWall Clock (s)∣ΔNLL∣\lvert\Delta\mathrm{NLL}\rvert ↓MMD* ↓SWD* ↓RDF W1 ↓Energy W1 ↓
FKC on MIT8,192585.79 ± 0.6414.2786 ± 0.14640.0200 ± 0.00090.0212 ± 0.00250.0424 ± 0.000914.2786 ± 0.1464
FKC on MIT20,0001423.86 ± 1.2314.1797 ± 0.01840.0180 ± 0.00040.0185 ± 0.00200.0415 ± 0.000714.1797 ± 0.0184
EBFKC on EAT (Ours)8,1921339.71 ± 0.450.9211 ± 0.39940.0600 ± 0.00590.0647 ± 0.00980.0121 ± 0.00750.9411 ± 0.3991

Under matched compute budgets, FKC on MIT still suffers from both sources of error and shows substantial sampling error. This agrees with our theory: naively adding particles cannot eliminate the energy's training error, so it does not help with accurate sampling from the target distribution.


We thank you again for your constructive feedback. We hope our responses and additional theoretical and experimental results have addressed your concerns, and we would be very grateful if you could consider raising your score.


Reference

[1] Skreta, Marta, et al. "Feynman-kac correctors in diffusion: Annealing, guidance, and product of experts." arXiv preprint arXiv:2503.02819 (2025).

[2] Meyer, Raphael A., et al. "Hutch++: Optimal stochastic trace estimation." Symposium on Simplicity in Algorithms (SOSA). Society for Industrial and Applied Mathematics, 2021.

Reviewer_S8wqOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the task of sampling for diffusion models with imperfect energy function instead of score function. Existing methods can suffer from endpoint mismatch and path-tracking errors. The authors propose an energy based Feynman–Kac corrector method. The method uses the pretrained energy to correct path-tracking error through Feynman–Kac reweighting, and aligns the terminal distribution with the reference energy. Experiments on Gaussian mixtures, interacting particle systems, ALDP and Ala-4 show the effectiveness of this method

Critical strengths

This paper studies an important task of sampling with inexact energy and presents an interesting formulation of correction method. Extensive molecular experiments provide strong support for the method.

Critical weaknesses

The main theoretical results concern the continuous-time case, but there is no analysis for more practical case such as discrete time sampler.

The two sources of error are discussed qualitatively, but no explicit quantitative bounds is included.

It is unclear how the method applies to standard score-based diffusion models

Questions for authors

How would the method apply to score-based diffusion models without explicit energy functions?

3 comments
AuthorsOpenReview ↗

Response to Reviewer S8wq [1/2]

We sincerely thank the reviewer for the careful review and constructive feedback. We are glad that the reviewer finds that our paper "studies an important task of sampling with inexact energy and presents an interesting formulation of correction method," and that "extensive molecular experiments provide strong support for the method." We address the comments and questions point by point below.

Quantitative error bounds

"The two sources of error are discussed qualitatively, but no explicit quantitative bounds is included."

We fully agree that both sources of error should be quantified explicitly. We now provide a bound for FKC on MIT, and the full derivation is included in the Appendix B.6 of the revised paper.

Notation. Let Et⋆=−log⁡pt⋆E_t^\star=-\log p_t^\star be the true intermediate energy, with Eref=ET⋆E^{\mathrm{ref}}=E_T^\star. We write Eθ(x,t)=Et⋆(x)+ct+δt(x)E_\theta(x,t)=E_t^\star(x)+c_t+\delta_t(x), where ctc_t is independent of xx and δt\delta_t is the energy approximation error. With vt⋆v_t^\star the true reverse drift and σt\sigma_t the diffusion coefficient, we define

\qquad \varepsilon_{\mathrm{end}} := \inf_{b\in\mathbb R}\lVert E^{\mathrm{ref}}-E_\theta(\cdot,T)-b\rVert_\infty \le \lVert\delta_T\rVert_\infty.$$ Let $q_t^{\mathrm{MIT}}\propto e^{-\gamma E_\theta(\cdot,t)+r_t}$ be the model-induced path and $q_T^{\mathrm{ref}}\propto e^{-\gamma E^{\mathrm{ref}}+r_T}$ the reference target. Let $\eta_T^{\mathrm{FKC}}$ be the terminal marginal of FKC in the continuous-time population limit, initialized from $q_0^{\mathrm{MIT}}$. **Result.** Under sufficient smoothness and integrability, $$d_{\mathrm{TV}}\left(\eta_T^{\mathrm{FKC}},q_T^{\mathrm{ref}}\right) \le \min\left\lbrace 1,\\, \underbrace{\frac{\gamma}{2}\int_0^T B_t\\,\mathrm{d}t}_{\text{path-tracking}} + \underbrace{\frac{\gamma}{2}\varepsilon_{\mathrm{end}}}_{\text{endpoint mismatch}} \right\rbrace.$$ Both terms are independent of the number of particles $N$ and the time discretization size $h$. EBFKC removes the path-tracking term by tracking the prescribed path exactly, and EAT removes the endpoint-mismatch term by pinning the terminal distribution to $q_T^{\mathrm{ref}}$. ### Discrete-time formulation > "The main theoretical results concern the continuous-time case, but there is no analysis for more practical case such as discrete time sampler." We formulate EBFKC in continuous time in Eq. (9) (lines 235–240) to provide a general framework for deriving and analyzing EBFKC sampling dynamics. The practical discrete-time sampler is obtained by discretizing both the particle dynamics and the log-weight updates. Specifically, Algorithm 1 uses Euler–Maruyama updates for the particles and left-point Euler updates for the log-weights, with the noise schedule of App. C.6, and the framework equally accommodates other time grids and consistent discretization schemes. This is exactly the Euler–Maruyama discretization, with resampling, of the augmented SDE $(X_t, A_t)$, where $\mathrm dA_t = G_t(X_t)\mathrm dt$ is the log-weight. Under standard smoothness and boundedness assumptions, the first-order weak convergence of the Euler scheme [1] gives an $O(h)$ discretization error, and resampling adds the standard $O(N^{-1/2})$ particle error [2]. Combined with our new bound above, the sampling error of FKC on MIT therefore comes from (i) path-tracking error and (ii) endpoint mismatch, which are independent of $h$ and $N$, together with (iii) an $O(h)$ discretization error and (iv) an $O(N^{-1/2})$ Monte Carlo error. Terms (iii) and (iv) vanish as $h\to0$ and $N\to\infty$ and are shared by all particle methods we compare, including FKC and EBFKC. Terms (i) and (ii), in contrast, are specific to the imperfect model and are exactly what EBFKC and EAT remove. We have added a remark in Appendix B.6.
AuthorsOpenReview ↗

Response to Reviewer S8wq [2/2]

Applicability to score-based diffusion models

"It is unclear how the method applies to standard score-based diffusion models. [...] How would the method apply to score-based diffusion models without explicit energy functions?"

We discuss this choice in Section 3.3. EBFKC's reweighting potential GtG_t (Theorem 2) involves ∂tΦt\partial_t\Phi_t, which requires the energy itself, not only its gradient. An energy-based model provides this, so EBFKC tracks the path defined by EθE_\theta exactly, even when EθE_\theta is imperfect. FKC instead implicitly follows a fictitious density associated with the score, and it tracks that density correctly only if the score is perfect.

For score-based models, there are two options:

  1. Recover the energy by integrating the score. This is costly and adds numerical error.
  2. Never resample at intermediate times. Itô's formula then telescopes ∂tΦt\partial_t\Phi_t into the endpoint potentials, and the weights depend on the model only through the score. For the EAT path, the terminal potential γEref−r\gamma E^{\mathrm{ref}} - r is model-free.

The second option, however, reduces to importance sampling on path space. It gives up both resampling and VCG, which itself requires ∂tΦt\partial_t\Phi_t, and both are key to avoiding weight degeneracy in our experiments. This is why we adopt energy-based modeling.

Importantly, the two failure modes we identify are not specific to energy-based models: endpoint mismatch and path-tracking error arise for any imperfect pretrained diffusion model, including score-based ones. Only their correction requires an energy, and obtaining one is not restrictive in practice. Energy-based diffusion models are an increasingly active direction, with recent training methods [3-5] and uses of the energy for composition and control with Monte Carlo methods [6-7]. Moreover, an existing score-based model can be distilled into an energy-based one by regressing −∇xEθ-\nabla_{\mathbf x}E_\theta onto its score [7]; the resulting energy supports SMC reweighting, so EBFKC can then be applied.


We thank you again for your constructive feedback. We hope our responses and additional results have addressed your concerns, and we would be very grateful if you could consider raising your score.


Reference

[1] Kloeden, Peter E., and Eckhard Platen. "Introduction to stochastic time discrete approximation." Numerical Solution of Stochastic Differential Equations. Berlin, Heidelberg: Springer Berlin Heidelberg, 1992. 305-337.

[2] Del Moral, Pierre. Feynman-Kac formulae: genealogical and interacting particle systems with applications. Springer New York, 2004.

[3] OuYang, RuiKang, Louis Grenioux, and José Miguel Hernández-Lobato. "A diffusive classification loss for learning energy-based generative models." arXiv preprint arXiv:2601.21025 (2026).

[4] Yu, Hanlin, et al. "Learning Energy-Based Models from Stochastic Interpolants using Spatiotemporal Differences." arXiv preprint arXiv:2605.26850 (2026).

[5] Guth, Florentin, Zahra Kadkhodaie, and Eero Simoncelli. "Learning normalized image densities via dual score matching." Advances in Neural Information Processing Systems 38 (2026): 89796-89826.

[6] Du, Yilun, et al. "Reduce, reuse, recycle: Compositional generation with energy-based diffusion models and mcmc." International conference on machine learning. PMLR, 2023.

[7] Thornton, James, et al. "Composition and control with distilled energy diffusion models and sequential monte carlo." arXiv preprint arXiv:2502.12786 (2025).

AuthorsOpenReview ↗

Follow-up on Response to Reviewer S8wq

As the discussion phase is approaching its end, we understand you are very busy, but we would greatly appreciate it if you could take a moment to review our response and let us know whether your concerns have been addressed. We would be very grateful if you could consider raising your score in light of our response. Thank you very much for your time.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to All Reviewers

We sincerely thank all three reviewers, S8wq, aQ3B, and idu7, for their time, careful reading, and constructive feedback.

We are encouraged that the reviewers find the problem we study important and meaningful (S8wq, aQ3B), our correction formulation interesting and useful (S8wq, aQ3B, idu7), and our empirical improvements substantial and well supported, from controlled Gaussian mixture experiments to learned molecular energy models (S8wq, aQ3B, idu7).

We are particularly pleased that all reviewers read the paper accurately and identified its key contribution. Existing inference-time scaling methods implicitly assume the pretrained diffusion model is exact. An imperfect model introduces two distinct errors, endpoint mismatch and path-tracking error, and adding more particles cannot remove them. Each summary captures this precisely. Reviewer S8wq notes that existing methods “suffer from endpoint mismatch and path-tracking errors.” Reviewer aQ3B notes that “the reference energy is incorporated into the prescribed path and reweighting to correct model error.” Reviewer idu7 notes that the paper “distinguishes two sources of error.” In particular:

  • S8wq finds the task “important” and the formulation “interesting,” and finds that “extensive molecular experiments provide strong support for the method.”
  • aQ3B finds that we address “a meaningful limitation of inference-time scaling,” and that “the controlled GMM experiments are particularly informative.”
  • idu7 finds that the paper “correctly distinguishes the density defined by a learned energy from the marginal distribution generated by the corresponding reverse SDE,” and that “the reported improvements are substantial across several relevant benchmarks.” The reviewer also rates the presentation as clear and easy to follow.

Summary of new material in this rebuttal

Following the reviewers' suggestions, this rebuttal provides:

  1. Revised title and framing (aQ3B). The new title is "Error-Corrected Inference-Time Scaling for Imperfect Energy-Based Diffusion Models." EBFKC is now framed as removing FKC's simplifying assumption.

  2. Quantitative error bound (S8wq). A total-variation bound on the sampling error of FKC on MIT that separates a path-tracking term and an endpoint-mismatch term. Neither term can be eliminated by increasing the number of particles or refining the time discretization.

  3. Computational complexity analysis (aQ3B). EBFKC on EAT costs O(NKRC)O(NKRC) versus O(NKC)O(NKC) for FKC, where CC is the cost of one energy gradient and RR the number of Hutchinson probes, so the relative overhead depends on system size mainly through RR. Empirically, EBFKC and FKC have nearly the same runtime on the same path, and evaluating the reference energy along the EAT path adds 1.1×1.1\times to 3.6×3.6\times in our experiments.

  4. FKC on MIT as a quantitative baseline (aQ3B). Results on GMM, DW-4, and LJ-13, under both annealing and reward tilting.

  5. Matched-compute comparisons (aQ3B, idu7):

    • FKC on MIT with more particles on LJ-13.
    • Geometric tempering SMC scaled up to 1,000,000 particles on LJ-13.
    • Particle-count scaling of EBFKC on EAT versus EBFKC on MIT on ALDP.

    In all cases, naively adding particles to a baseline does not close the gap, while EBFKC on EAT benefits from additional particles.

  6. Scaling study on probes count and runtime (aQ3B): Experiments on coupled Gaussian systems show that probes count and runtime remain manageable as system dimension increases.

  7. Evaluation on hexa-alanine (aQ3B): Results on EBFKC's application to hexa-alanine annealing.

  8. Clarifications:

    • the discrete-time sampler (S8wq);
    • applicability to score-based models and to settings without a reference energy (S8wq, aQ3B);
    • the differences between EBFKC and DABG (idu7).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.