acceptodds
Under review as a conference paper at ICLR 2027

Kalman Posterior Sampling with Diffusion Priors and Simulation-Based Likelihoods

Abstract

In science, measurements are often modeled by stochastic numerical simulators whose randomness is non-Gaussian and whose likelihood and gradients are rarely available. Diffusion posterior sampling usually assumes all three, which rules these simulators out. We introduce Kalman Posterior Sampling (KPS), a sampler in which the diffusion prior and a statistical linearization of the likelihood defined by the simulator enter a common Kalman update through their moments, using simulations alone or gradients when the simulator is differentiable. An inner Gibbs loop keeps the iterates consistent with the diffusion process, preventing errors from accumulating across steps. On latent fluid dynamics and scientific inverse problem benchmarks, KPS outperforms state-of-the-art samplers in accuracy with as few as 16 particles, delivers close-to-calibrated uncertainty, and captures multimodal posteriors even from black-box simulators.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Bayesian inference, posterior sampling, inverse problems, generative priors, Kalman

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_UqGnOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Kalman Posterior Sampling (KPS) is a method for diffusion posterior sampling that addresses a more general form of inverse problems in which the measurement likelihood is intractable. KPS decomposes posterior sampling into transitions p(xt−1∣xt,y)p(x_{t-1} \mid x_t, y). It approximates p(x∣xt)p(x \mid x_t) using Tweedie identities and uses statistical linearization to form a Gaussian approximation to p(y∣x)p(y \mid x); this yields an algorithm that iteratively denoises samples and contains an inner Gibbs loop to apply Kalman updates and renoising at each noise level. The paper presents a gradient-based variant and a more general simulator gradient-free variant, demonstrating strong results across several scientific inverse problems.

Critical strengths

1. The problem is well-motivated and timely. Derivative-free posterior sampling for inverse problems with inaccessible likelihoods is practically useful for real world scientific tasks. The focus on evaluation of posterior coverage, rather than simply on deterministic reconstruction metrics, is valuable.

2. The proposed algorithm is conceptually simple and well motivated: it combines diffusion-based local Gaussian approximations with statistical linearization and a Kalman update, while the re-noising/Gibbs loop provides a principled way to maintain consistency across diffusion levels.

Critical weaknesses

1. I'm concerned about the novelty and contribution of KPS relative to existing work like Blade [1], which also focuses on derivative-free posterior sampling and uses ensemble statistical linearization in a Gibbs-style sampler. The paper briefly mentions Blade in the related work but does not include comparison against it in the experiments, despite Blade, to my understanding, being the closest prior work to KPS. This is especially important for Navier–Stokes, where Blade evaluates on the same InverseBench problem (as far as I can tell) and, based on the published results, substantially outperforms KPS-H at the same ensemble size.

2. I don't think the experiments convincingly validate the generality of KPS to problems with intractable measurement likelihoods. Several of the evaluated tasks use the standard additive Gaussian measurement noise; while the black hole imaging task does not, the results are not fairly comparable with the baselines due to the difference in noise model as reported by the paper. So the experiment most relevant to the paper’s central motivation lacks a controlled comparison to prior methods under the same observation model.

Questions for authors
  1. Major: can the authors clarify the novelty/contribution of KPS w.r.t. Blade? And can they please provide experimental comparison with Blade, such as on Navier Stokes and the other tasks?
  2. Moderate: Black hole imaging appears to be the main experiment that directly tests the non-additive/non-Gaussian, simulation-based likelihood setting. But there isn't a fair comparison to the baselines due to the difference in noise model. Could the authors adapt some representative baselines to the same noise model setting to enable fair comparison? It doesn't have to be all the reported baselines, just a few could be useful. This would help emphasize the strength of KPS's general likelihood-free formulation.
  3. Clarification: The paper provides per-update complexity and simulator-call counts, but no end-to-end runtime or matched-compute comparison against baselines. Could you report wall-clock time and/or total simulator/denoiser evaluations for KPS-H and KPS-G relative to the closest baselines, especially Blade on Navier–Stokes?

[1] Zheng et al., TMLR “Blade: A Derivative-free Bayesian Inversion Method using Diffusion Priors.”

Most concerns have been satisfied. Updating score from 2 to 3.

5 comments
AuthorsOpenReview ↗

Direct answer

We thank the reviewer for insisting on BLADE and for the concrete suggestions.

W1 & Q1: Novelty with respect to BLADE, and experimental comparison.

We agree that BLADE is the closest prior work and that the submission did not acknowledge it enough. Both methods build on the ensemble statistical linearization of EnKG. They differ in how they use it and in what they assume about the likelihood:

  1. Observation model. KPS never uses the clean observation operator. It only queries the stochastic simulator, so the innovation (y−y(i))(y - y^{(i)}) carries a draw of the observation noise and no noise model has to be specified. BLADE evaluates the clean operator.
  2. Update. BLADE uses the linearization inside a likelihood step made of many small Langevin moves preconditioned by the ensemble covariance. KPS uses it in a single closed-form Kalman update of p(x∣xt)p(x \mid x_t), whose step is set by the gain at the cost of a linear solve.
  3. Local prior. The KPS update combines the likelihood surrogate with a per-particle prior N(μt,Vt)\mathcal{N}(\mu_t, V_t) whose covariance comes from the denoiser through Tweedie's formula, rather than with the covariance of the cloud.
  4. Tuning. BLADE requires a step size and an observation covariance tuned for calibration. KPS estimates Σy\Sigma_y from the held-out residual of the same regression that gives AA, at no extra simulator cost (Appendix A.6), and its update has no tuned parameters, only budget knobs.
  5. Gradient variant. KPS-G replaces the regression slope by the expected Jacobian. Stein's lemma connects the two (Section 4.2). With gradients, KPS-G reaches our best accuracy and calibration with 16 particles instead of 512.

These choices have a cost. A single closed-form step inherits the error of the linear surrogate when the forward model is strongly non-linear, where BLADE's small steps are likely more robust. We see the two methods as complementary. For example, BLADE could query the simulator and estimate Σy\Sigma_y as KPS does, and KPS could relinearize within a noise level. We now describe BLADE in detail in Section 3 and discuss the comparison in a new paragraph, "Relation to BLADE" (Section 6).

On experiments, we added BLADE's reported Navier–Stokes results (N=512N = 512, most calibrated configuration) to Table 2. The reviewer is right: BLADE outperforms KPS-H there in both relative L2L_2 (0.110 vs. 0.313) and calibration. We attribute this to the strong non-linearity of the PDE solver, where a single Kalman step on a linear surrogate is less robust than many small steps. We are also adding BLADE to the latent problem, where all methods are run by us under identical settings, and will share these results during the discussion.

W2 & Q2: Controlled comparison on black-hole imaging.

Rather than adapting the baselines to our noise model, we adapted KPS to theirs. Both variants now also run under the eht noise model used for the published baselines (Table 2, Table 8, Appendix C.3.1). Under this model, KPS queries the simulator directly and both variants still exceed the best published blur PSNR (34.8 and 34.7 vs. 32.9). The eht model applies non-differentiable station gains, so KPS-G must take its Jacobian from the differentiable vis_thermal model. This makes it a misspecified-gradient setting, in which KPS-G remains close to calibrated (MIRA lower by only 0.028±0.0060.028 \pm 0.006, Fig. 9). Figure 8 quantifies how the two noise models differ.

On the broader point, the other tasks have tractable likelihoods by design: the baselines need an explicit likelihood to be run at all. KPS never uses it and treats every task as if the likelihood were intractable. We will also add a non-Gaussian (Laplace) noise experiment on the latent problem.

Q3: End-to-end cost.

We will report wall-clock time and peak memory for all methods, including BLADE, on the latent problem, normalized per posterior sample since KPS produces the whole ensemble in parallel. No method, including ours, was tuned for efficiency: we used the configurations of the original code.

While waiting for additional results on BLADE on the latent problem, we do believe that our responses have addressed your main concerns. We would be happy to provide further clarification if any issues remain. Otherwise, we would be grateful if you could reconsider your score in light of our revisions.

Thanks for the response. The clarification of the methodological differences between KPS and BLADE is appreciated. However, the Navier–Stokes results show that BLADE substantially outperforms KPS-H in both reconstruction accuracy and calibration at the same ensemble size. This leaves the practical benefit of KPS unclear.

The additional black hole experiments under the baselines’ observation model are useful and an improvement.

My question about end-to-end computational cost remains unresolved. The reported operation counts alone do not establish the computational tradeoff relative to the closest baselines.

I appreciate the improvements, but the remaining uncertainty about the practical contribution and computational cost leads me to maintain my weak reject score.

AuthorsOpenReview ↗

BLADE on latent problem and further results

We thank the reviewer for the follow-up and for acknowledging the black-hole experiments. We now have first results with BLADE on the latent problem, which bear directly on both remaining concerns.

Navier–Stokes. Beyond the reasons discussed in the revised paper, BLADE's results on this task rely on careful per-task tuning. Its hyperparameters are set per task and per noise level (BLADE, Table 4), with the effective observation noise σ~y\tilde\sigma_y ranging from 0.03 to 0.3. The BLADE authors identify σ~y\tilde\sigma_y as the most important hyperparameter ("Impact of different hyperparameters", Section 4.3). They set it below the true noise level and tune it so that the spread-skill ratio is close to 1 ("Effective observation noise", Appendix B.1). Its calibration on Navier–Stokes is therefore partly obtained by tuning toward the calibration target. KPS has no step size or observation noise to tune. It estimates Σy\Sigma_y from the simulated pairs, and uses a single configuration across the InverseBench tasks (Table 5).

BLADE on the latent problem. We adapted the official implementation (blade.py) to our setting. We used its main (most calibrated) version with 2048 particles, as for all ensemble methods except KPS-G, and 60 Langevin steps. Following Appendix B.1 of the BLADE paper, we set σ~y\tilde\sigma_y below the true observation noise σy=0.1\sigma_y = 0.1. We kept the ratio of their Navier–Stokes configuration at σnoise=1\sigma_{\text{noise}} = 1 (σ~y=0.17\tilde\sigma_y = 0.17), which gives σ~y=0.017\tilde\sigma_y = 0.017, without further tuning. The results are in the updated Fig. 3:

MethodSkill ↓SSR → 1CRPS ↓
BLADE0.32 ± 0.130.38 ± 0.09425 ± 214
KPS-G0.23 ± 0.031.17 ± 0.12266 ± 39
KPS-H0.23 ± 0.031.04 ± 0.10266 ± 36

On this task, BLADE is under-dispersed and varies strongly across instances. The galleries show why. On some instances, BLADE is qualitatively on par with the best methods (Figs. 19 and 20). On others, it does not diverge but misses the target entirely (Figs. 21 and 22). This is consistent with the BLADE paper's finding that σ~y\tilde\sigma_y must be tuned per problem: a setting that works well on one task does not transfer directly to another.

Black-hole imaging. We also realized that BLADE reports black-hole imaging results (BLADE, Appendix Table 3), for point accuracy only, and we have added them to Table 8. Under the same eht noise model, BLADE (main) leads in PSNR (31.3 vs. 28.9 for KPS-G) and in closure-phase fit, while KPS-G fits closure amplitudes better (χcamp2\chi^2_{\text{camp}} 1.57 vs. 1.95). BLADE reports no calibration metrics on this task, whereas KPS-G remains close to calibrated under both noise models (Figs. 4 and 9), with the same configuration in both.

Computational cost. With KK particles, TT noise levels (or split-Gibbs iterations), NN Langevin steps and MM Gibbs sweeps, the number of forward-model calls scales as follows:

MethodForward-model callsSimulator gradient
DPS, MMPS, TDSO(KT)O(KT)yes
EnKGO(KT)O(KT)no
DAPS, PnP-DMO(KTN)O(KTN)yes
BLADEO(KTN)O(KTN)no
KPS-HO(KTM)O(KTM)no
KPS-GO(KTM)O(KTM)yes

Concretely, BLADE's Navier–Stokes configuration runs 25 split-Gibbs iterations of 50 Langevin steps each (BLADE, Appendix Table 2). This amounts to 1,250 forward-model calls per particle, against 64 for KPS (T=32T = 32, M=2M = 2). At 512 particles, BLADE makes 640,000 calls against 32,768 for KPS-H, about 20 times more. This matters when the simulator is expensive. We have added this table to Appendix A.8.

Our preliminary observations on wall-clock time are the following:

  • In these benchmarks, the simulator is cheap, so calling it less often does not yet show in wall-clock time.
  • The main costs of KPS are the linear solve of the Kalman update and the vector-Jacobian products through the denoiser for the Tweedie covariance. We keep both because they make the update fully defined, with no step size or noise level to tune.
  • KPS-G is particle-efficient. Its slope has full rank whatever the ensemble size, so few particles suffice: 16 on InverseBench, and 64 on the latent problem against 2048 for the other ensemble methods.
  • Neither variant has been optimized for speed, and standard engineering could reduce their cost.

We ran into implementation issues while setting up the timing, memory and Laplace noise experiments, and are working to resolve them. We will post these results as soon as they are available, and aim to do so before the end of the discussion period. We hope the results above already address the reviewer's main concerns, and would be grateful if they could reconsider their score in their light, or let us know what would help them do so.

Reviewer_1XETOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper develops a Kalman filtering-inspired method to perform diffusion posterior sampling, where the goal is to sample from the posterior defined by a prior diffusion model and an observational model. Crucially, this observational model, like in many scientific problems, is a stochastic forward simulator and does not emit a likelihood, and sometimes can be non-differentiable. While most prior methods consider tractable or Gaussian likelihood and differentiability, the paper instead uses a Kalman-style update to approximate each observation-conditioned denoising step. This is made possible by using a linear gaussian approximation to the forward model. Furthermore, to solve for the relevant parameters in Kalman update, the paper develops two schemes, one estimating the moments by simulating multiple samples from the forward model and the other making using of the gradient, if differentiable, through the Gaussian stein identity. The proposed methods are evaluated on a semi-synthetic latent diffusion model task and InverseBench. Overall, the gradient-based approach seems to have the best performance in most tasks.

Critical strengths
  • The use of Kalman update is diffusion posterior sampling is novel.

  • The proposed moment-matching method does not require differentiable forward model, which can be applied to a range of scientific problems.

  • The derivation of the Kalman update is intuitive and easy to follow.

Critical weaknesses

I found the developed method contradicts the motivation to deal with scientific measurements that "are often modeled by stochastic numerical simulators whose randomness is non-Gaussian and whose likelihood and gradients are rarely available."

  • The Kalman update is derived based on linear gaussian model -- that's assuming certain structure on the forward simulator which may not always hold. For example, the forward model can be multi-modal or have non-Gaussian noise. It's unclear whether Kalman update is accurate or not for these cases.
  • The gradient-based variant makes use of Stein Gaussian identity, which assume the data x is gaussian distributed.
  • While the gradient-based variant seems to be the best overall performance, it does not align with the motivation on solving non-diffenrentiability.

For the moment-matching method, while working in more general regime, I have concerns on its computational cost as well as the approximation accuracy in high dimensions, since at each sampling step multiple samples need to be drew.

Moreover, there is no approximation error quantification or discussion of the proposed method.

Finally, it's unclear whether the empirical comparison is done at matched compute. So it makes it challenging to access the actual performance gain of the proposed method.

Questions for authors
  • Can the authors explain Figure 1, in particular, what are the blue dots, grey dots, contours and the curve with the uncertainty band?

  • Do the authors compute VtV_t in eq(3) exactly? In that case, wouldn't it be compute-extensive as it requires computing the Jacobian of the score network?

  • Can the authors explain the M-step loop in Algorithm 1? I see that it's emulating the Gibbs loop; however, as the Kalman update is imperfect, would running more steps always improve the performance? Have the authors considered running any ablation on M?

  • Have the authors considered a matched compute comparison to other baselines, or report the compute cost of each method for the current results in the paper?

  • Can the authors explain the metrics used in Figure 3 in more details, in particular, whether higher the better or the reverse? In addition, TARP and MIRA are heavily references in the paper so please consider providing some relevant background.

3 comments
AuthorsOpenReview ↗

Direct answer - Weaknesses

We thank the reviewer for the detailed review and the precise questions.

W1: Linear-Gaussian assumption vs. the motivation.

The Kalman update needs a linear-Gaussian surrogate, not a linear-Gaussian simulator. The surrogate is fitted from simulated pairs by statistical linear regression, which matches the first two moments of (x,y)(x, y) whatever the simulator and its noise. Σy\Sigma_y absorbs both the simulator noise and the misfit of the affine fit. This remains an approximation, and we agree the paper stated its consequences too quickly. Section 4.2 now makes them explicit: treating the joint as Gaussian holds only for a Gaussian cloud pushed through an affine simulator with Gaussian noise, and otherwise the surrogate misses the curvature of the simulator and the shape of its noise.

On multimodality, each local posterior is Gaussian, but each particle is updated from its own anchor. The cloud is therefore a mixture, and the inner Gibbs loop lets particles move between modes (Section 4.3). Figure 5 shows two distinct modes recovered in one black-hole run. On non-Gaussian noise, black-hole imaging is non-Gaussian in image space, and we added runs under the eht noise model with station gain errors (Appendix C.3.1). We will also add a Laplace noise experiment on the latent problem. Natural extensions, such as Gaussian-mixture likelihoods fitted on the simulated pairs, are now discussed in "Limitations and extensions" (Section 6).

W2: Stein's identity.

We agree, and we now state it explicitly in Section 4.2. KPS-G does not rely on Stein's identity to be valid. Its slope is the expected Jacobian, a linearization in its own right. Stein's lemma only shows that it coincides with the KPS-H slope when the cloud is Gaussian. This is an additional assumption, distinct from the Gaussian local prior of Eq. 13: each p(x∣xt(i))p(x \mid x_t^{(i)}) is Gaussian, but their mixture over anchors is not. The two slopes therefore differ in general, and coincide whatever the cloud only for a simulator affine in xx.

W3: The gradient variant and non-differentiability.

Our aim is not to solve non-differentiability alone, but to provide one framework that uses whatever the problem offers: simulations only, or gradients when available. Gradients matter in practice. With them, KPS-G reaches our best accuracy and calibration with 16 particles instead of 512 (Table 2), which matters when simulations are expensive. KPS-G also shares the ingredients of MMPS (Tweedie covariance, Jacobian-based gain) but only queries the stochastic simulator and estimates Σy\Sigma_y from the pairs, with no tuning. Gradients also need not come from the exact simulator. On black-hole imaging, KPS-G with its Jacobian taken from a different simulator than the observations remains close to calibrated (Section 5.2, Appendix C.3.1).

W4, W6 & Q4: Computational cost and matched compute.

The per-particle cost of one update is given in Table 3 (Appendix A.8): one simulation per particle and update for both variants, plus derivative products whose number we report. Matching compute exactly is difficult, since the methods use different resources: simulations, simulator gradients, denoiser evaluations or Langevin steps. We will therefore report wall-clock time and peak memory for all methods on the latent problem, normalized per posterior sample since KPS produces the whole ensemble in parallel. No method, including ours, was tuned for efficiency: we used the configurations of the original code.

On accuracy in high dimension, the InverseBench states have up to 16,384 dimensions. KPS-H degrades there because its slope has rank at most min⁡(r,N−1)\min(r, N-1), and it improves with NN (Table 9, Figs. 10 and 11). KPS-G is unaffected, as its slope is full rank whatever the ensemble size.

W5: Approximation error.

KPS is approximate by design, and we do not claim a bound on the final error. Appendix A.7 analyses what we can: the inner Gibbs loop has the target as its unique stationary distribution when the update is exact. With our surrogate, the error at a given level splits into a mixing term that decays with MM and the bias of the surrogate. The re-noise step never amplifies the error it receives (Eq. 43). We now make each approximation and its consequences explicit (Section 4.2, "Limitations and extensions" in Section 6). Asymptotically exact samplers such as TDS avoid this bias but need a likelihood density, which a simulator does not provide. Empirically, we quantify the error through posterior-level calibration (TARP, MIRA) and a posterior predictive check that needs no ground truth (Appendix C.4).

AuthorsOpenReview ↗

Direct answer - Questions

Q1: Figure 1.

Contours show the prior p(x)p(x). The black curve with its grey band shows the states consistent with the observation, where p(y∣x)p(y \mid x) has mass. Dots are particles, in blue after each step and in grey before. Panels 1 to 4 show one noise level: noisy particles xtx_t (1) are denoised into draws from p(x∣xt)p(x \mid x_t) (2), moved toward the band by the Kalman update (3), and re-noised for the next sweep or noise level (4). The right panels show the band as seen by the update, linearized through AA, which is fitted once over the ensemble (KPS-H) or computed from each particle's Jacobian (KPS-G). We have rewritten the caption accordingly.

Q2: Jacobian of the denoiser.

We assume the reviewer refers to the covariance VtV_t of Eq. 13, which involves the Jacobian of the denoiser. Eq. 3 is the likelihood, which KPS never evaluates. VtV_t is never formed: only vector-Jacobian products through the denoiser are needed (Appendix A.2), as in MMPS. KPS-H needs rr of them per particle, after which the gain reduces to an r×rr \times r solve. KPS-G needs one per GMRES iteration (Appendix A.5, Table 3).

Q3: The MM loop in Algorithm 1.

More sweeps do not always help, for the reason the reviewer gives. Following Rochman-Sharabi & Louppe (2026) [PPR], the error of an approximate corrector splits into a mixing term that decays with MM and the bias of the inner update, which sweeps cannot remove (Appendix A.7). Gains therefore saturate, while the cost grows linearly in MM. We use M=2M = 2 throughout (Table 5).

Q5: Metrics.

We added a description of all metrics in Appendix B. In Fig. 3, Skill (RMSE of the ensemble mean) and CRPS are better when lower. The spread-to-skill ratio should be close to 1: below 1 means over-confident, above 1 under-confident. MIRA equals 2/3 for a calibrated posterior and 1/2 for one unrelated to the truth. TARP plots expected coverage against credibility level, and a calibrated posterior follows the diagonal. Appendix B also explains how over- and under-confidence show in TARP and MIRA.

[PPR] Predict-Project-Renoise: Sampling Diffusion Models under Hard Constraints

We believe that our responses have addressed your main concerns. We would be happy to provide further clarification if any issues remain. Otherwise, we would be grateful if you could reconsider your score in light of our revisions.

I would like to thank the authors for their rebuttal. Most of my concerns are resolved. However,I am still a bit skeptical about the effectiveness of using a Guasisan approximation for multimodal target / non-Gaussian target which is the primary motivation of the work. I'd also encourage rescoping the paper around the non-differnetiability part since there is a differentiable version of the algorithm. Nevertheless, this work bridges a gap in the literature and I would update my score to 3.

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

This work proposes the Kalman Posterior Sampler (KPS), a training-free sampling method that addresses the sampling of an intractable posterior p(x∣y)p(x|y) given a pre-trained diffusion model for the distribution of xx. KPS assumes that yy can be easily simulated from xx but p(y∣x)p(y|x) is intractable. It uses an ensemble of particles at each noise level (step) of the denoising process to estimate the intractable posterior distribution p(x∣y)p(x|y), where each particle is a candidate sample of xx. At each step, KPS denoises the particles, pushes them through the simulator, and fits an affine Gaussian surrogate of the likelihood from the resulting (x,y)(x, y) pairs. It then updates each particle using a closed-form Kalman update, with the prior covariance derived from the denoiser (Sec. 4.2). A re-noise step is applied multiple times to keep the particles consistent with the distribution induced by the forward diffusion process.

The authors propose and study two variants of KPS: KPS-Hybrid, which fits the slope of the affine likelihood surrogate y≈Ax+by \approx Ax + b via regression over the set of simulated particles, and KPS-Gradient, which relies on the simulator Jacobian. Experimental results on the latent-space fluid problem and InverseBench tasks demonstrate competitive performance against existing baseline methods. On the InverseBench tasks, KPS-Gradient achieves the best point accuracy with 16 particles and is close to being well-calibrated, while KPS-Hybrid is weaker and under-dispersed.

Critical strengths
  1. The problem is precisely framed in the introduction, providing clear motivation and establishing the context for the proposed KPS method.

  2. The method is presented clearly and demonstrates careful design. The paper shows that only one step per noise level requires approximation—inferring the clean x from the current noisy particle and the measurement—while the rest of the chain remains exact. This step is handled using a standard Kalman update, drawing on readily available components like the denoiser and simulated pairs. This is followed by safeguards to prevent ensemble collapse and ensure the particles remain consistent with the forward diffusion process.

  3. The experimental settings and evaluations are equally thorough. The authors assess whether the uncertainty is well-calibrated, rather than just checking reconstruction accuracy. For the latent fluid task, they evaluate spread-skill and MIRA across all methods. On InverseBench, they report TARP, MIRA, rank histograms, and coverage for the KPS variants. The experiments also compare KPS-Gradient and KPS-Hybrid. KPS-Gradient outperforms the best published results using just 16 samples, while the authors candidly acknowledge that KPS-Hybrid is weaker and under-dispersed.

Critical weaknesses

I have a few minor concerns about the work:

  1. KPS-Hybrid, which offers the gradient-free property that KPS-Gradient lacks, underperforms KPS-Gradient. On inverse-scattering and Navier–Stokes problems, it also underperforms the other baseline methods.
  2. Computational cost and variability due to stochasticity are not discussed or compared. Since KPS relies on maintaining an ensemble of particles and multiple steps of stochastic simulation, it is not clear to the reader how it compares against other methods in terms of computational cost (e.g., wall-clock time or FLOPS). Given the stochasticity of the simulation, it would also be beneficial to report confidence intervals across multiple seeds for the metrics in Fig. 3.
  3. It seems that the latent diffusion model is not within the target setting introduced in Section 1, where p(y|x) is intractable. Could the authors explicitly confirm this?
Questions for authors

Apart from the weakness, can the author provide some intuitive explanation on why underperforms KPS-Gradient and why KPS-Hybrid fails on the InverseBench problems like Inverse-scattering and Navier-Stokes but performs relatively well on the latent diffusion model problem? It's a great topic to be included in Sec. 6's discussion.

2 comments
AuthorsOpenReview ↗

Direct answer

We thank the reviewer for the careful reading and the constructive questions.

W1 & Q: KPS-Hybrid vs. KPS-Gradient and the baselines.

Whether KPS-H underperforms depends on the metric. On the latent problem, it matches KPS-G in skill (0.234 vs. 0.233) and CRPS (266.3) and is better calibrated (SSR 1.038, and a MIRA score within uncertainty of 2/3, Fig. 3). On black-hole imaging, it exceeds the best published blur PSNR under both noise models (36.9 and 34.8 vs. 32.9, Table 2). Point accuracy against a single truth is also not what a calibrated sampler optimizes. A calibrated posterior spreads its samples over all states that explain yy, which raises the error against x⋆x^\star. The published InverseBench baselines report only point accuracy, and Sharief et al. (2026) [MIRA] find that the methods doing well on it are not calibrated.

The comparison is also not controlled. The baselines evaluate an explicit likelihood built on the clean observation operator. KPS only queries the stochastic simulator and must estimate the whole likelihood from noisy samples. We now state this explicitly in Section 5.2.

The gap with KPS-G is nonetheless real, and we see two causes.

Rank. KPS-H fits one slope over the cloud, of rank at most min⁡(r,N−1)\min(r, N-1), whereas KPS-G uses full-rank Jacobians. We kept r=32r = 32 fixed across all tasks (Table 5) to avoid per-problem tuning. This truncation restricts the correction to a few directions, which costs more as the state dimension grows. Consistently, KPS-H does well on the latent (2,048 dimensions) and black-hole (4,096) problems and trails on the two 128×128128 \times 128 tasks (16,384). On inverse scattering, its posterior mean barely changes with NN while the spread grows (Fig. 11), as expected from a correction confined to a fixed number of directions.

Non-linearity. On Navier–Stokes, the forward model is a full PDE solve. A single Kalman step on a linear surrogate can then be overconfident where the surrogate is poor. Methods taking many small steps (BLADE) or targeting a MAP-like solution (EnKG at N=2048N = 2048) do better there. KPS-H still matches DPG, the second-best published result (0.313 vs. 0.325).

We have added a discussion of the assumptions behind the surrogate and their consequences (Section 4.2, and "Limitations and extensions" and "Relation to BLADE" in Section 6). We will add the rank explanation to "Choosing a variant", as the reviewer suggests.

W2: Computational cost and variability.

The per-particle cost of one update is given in Table 3 (Appendix A.8), but we agree that a direct comparison is missing. We will report wall-clock time and peak memory for all methods on the latent problem. These will be normalized per posterior sample, since KPS, like EnKG, produces the whole ensemble in parallel. No method, including ours, was tuned for efficiency: we used the configurations of the original code. We will also add ± standard deviations to the Skill, SSR and CRPS of Fig. 3 [across the 50 reference instances / across seeds]. The MIRA scores in Fig. 3 already have error bars, and the InverseBench results report mean ± std (Tables 6–8) or interquartile ranges (Table 9).

W3: Tractability of the latent problem.

We confirm that the latent likelihood is tractable (additive Gaussian noise after a known decoder), as are those of inverse scattering and Navier–Stokes. This is deliberate, since the baselines need an explicit likelihood to be run at all (Section 5.1). KPS, however, never uses it. It only queries the simulator, so it treats every task as if the likelihood were intractable.

The tasks form a progression from problems with known likelihood components to black-hole imaging, where the literature already has to approximate the likelihood [ref]. On that task, we added a misspecified setting in which KPS-G takes its Jacobian from a different simulator than the one generating the observations. KPS-G remains close to calibrated there (Section 5.2, Appendix C.3.1). We will also add a non-Gaussian (Laplace) noise experiment on the latent problem.

[MIRA] MIRA: A Score for Conditional Distribution Accuracy and Model Comparison

We believe that our responses have addressed your main concerns. We would be happy to provide further clarification if any issues remain. Otherwise, we would be grateful if you could reconsider your score in light of our revisions.

Reviewer_tiThOpenReview ↗

Post Rebuttal Comment

I would like to thank the authors for acknowledging the gap between KPS-H and KPS-G and making commitments for further evaluating and comparing the computational costs for future versions. The lack of full direct comparison does not fully address my concerns and similar concerns from other reviewers. I am unable to increase my score to a full accept and will maintain my current rating.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Global answer

We thank the reviewers for their constructive feedback. We will first restate the main message of the paper, which we hope addresses some of the concerns. We understand these concerns still deserve direct answers and summarize how we plan to address them.

Main message

KPS is not meant to be the most optimized sampler for a given task. It is a single framework whose two variants adapt to what the forward model provides (simulations only, or gradients when available). It has no hyperparameters beyond budget knobs (particles, noise levels, sweeps, rank or Krylov dimension) and no guidance scale or step size to tune. Our baselines rely on known parts of the likelihood: the noise model, the clean observation operator, or the density itself. KPS uses none of them and only queries the forward model. This puts it at a disadvantage, yet it remains competitive across a broad range of tasks.

Our second point concerns evaluation. To the best of our knowledge, few if any diffusion posterior sampling papers on such problems go beyond point accuracy or marginal calibration (Skill, CRPS, SSR). We add TARP and MIRA because we consider them necessary diagnostics in scientific settings, where calibration is what matters.

Beyond Gaussian noise

We chose tasks that span different settings, connect easily to prior work, and come with pre-trained priors. They form a progression from problems with known likelihood components to black-hole imaging, where even the literature must approximate the likelihood [1], which shows that realistic problems quickly become intractable. We will extend this task with a misspecified setting, in which the gradients come from a different simulator than the observations, to further illustrate the flexibility of KPS. To probe non-Gaussian noise in a controlled way, we will also add a Laplace noise experiment on the latent problem.

[1] How narrow is the M87* ring? I. The choice of closure likelihood function (link to the paper)

Assumptions and approximations

Our setting leaves the likelihood fully unknown to the method, and the Kalman update relies on a Gaussian local prior, which in turn constrains the shape of the local posterior. We will state each approximation explicitly along with its consequences. We will also make precise when Stein’s lemma makes KPS-H and KPS-G estimate the same slope. Finally, we will discuss extensions such as Gaussian-mixture likelihood surrogates, which stay within the framework at higher cost, local linearizations, which become more accurate as t -> 0, ...

BLADE

We agree that BLADE was not compared fairly. We will add it as a baseline and clarify in the related work how it differs from KPS. BLADE embeds the linearization in a split Gibbs sampler with a Langevin likelihood step. KPS uses it in a single closed-form Kalman update.

Planned changes

Below are the changes we commit to. Each reviewer is pointed to those addressing their concerns. We wanted to initiate the discussion quickly as several questions can already be addressed. A first update of the pdf will be uploaded along with answers. We will generate the latex diff to better guide reviewers to changes we did.

Latent problem

  • ± standard deviations for Skill, SSR and CRPS
  • a Laplace noise experiment
  • BLADE as a baseline
  • time and memory for all methods. No method, including ours, was tuned for efficiency, so the cost comparison is indicative.

Navier–Stokes

  • BLADE’s reported results added to Table 2

Black-hole imaging

  • we used vis_thermal noise, which makes the simulator differentiable, rather than the InverseBench default eht. We added runs of KPS-H (N = 512) with eht as a well-specified control for fair comparison. We also added KPS-G with vis_thermal gradients but eht simulations, a misspecified simulator setting that KPS supports at no extra cost.

Metrics

  • a complete definition of Skill, CRPS, SSR, TARP and MIRA in Appendix.

And, of course, adaptation of the main text to address reviewers' concerns. We will be more exhaustive in per-reviewer answers.

Correction

The CRPS values in Tables 2 and 9 were inadvertently reported normalised by the RMS of the ground truth, which is not the standard definition, is inconsistent with our latent experiment, and precluded direct comparison with published numbers. We have corrected all entries to the raw (unnormalised) fair CRPS.

AuthorsOpenReview ↗

End of discussion period

We thank all reviewers and the area chair once more for their time and for a discussion that substantially improved the paper. As the discussion period closes, we summarize our last changes and what we commit to for the final version.

Bug fix on the latent problem. While preparing the remaining experiments, we found a bug in our implementation for the latent problem (see our discussion with Reviewer UqGn). We corrected it and re-ran the affected experiments to make sure all results are consistent, and the updated results are now in Fig. 3. The overall picture is unchanged, but some numbers moved: KPS-Gradient is now much better calibrated than before, while KPS-Hybrid lags slightly behind it.

Non-Gaussian noise experiment. The bug fix delayed our additional experiments on the latent problem. We originally planned a Laplace noise model, but its effect on the observations may remain close to the unimodal Gaussian case. Unless the reviewers object, we will instead use a bimodal Gaussian mixture noise, which changes the structure of the observations more substantially. If a reviewer prefers the Laplace noise model, we will keep it.

Time and memory. We know the cost comparison is an expected result, and we chose to design it carefully rather than report it hastily. It will not be ready before the end of the discussion, but the final version will include:

  1. End-to-end time and memory tracking of all methods on the latent problem. Our empirical observations so far concerning kps are summarized in the discussion with Reviewer UqGn.
  2. A theoretical and empirical breakdown of the cost of each step of KPS, so that readers can identify the main sources of slowdown and where optimization pays off first. We acknowledge that solving a linear system and building the full surrogate (slope and covariance) has a cost and makes some parts of KPS slower.

Presentation. The rebuttal added many results, and we adapted the text accordingly. For the final version, we will restructure the paper into a clear and readable story within the 10-page limit, and integrate all results obtained during the discussion.

Thank you again for your engagement throughout the discussion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.