Pith. sign in

REVIEW 2 major objections 1 minor 13 references

Latent diffusion posterior sampling uses a neural surrogate to guide PDE inverse problems without repeated full-model evaluations.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

L-DPS performs approximate Bayesian inference for PDE inverse problems by combining latent diffusion priors with surrogate-guided likelihood in a lower-dimensional space, demonstrated on Darcy flow permeability inversion.

T0 review reviewed 2026-06-26 challenge →

load-bearing objection L-DPS assembles VAE latent diffusion with DPS and a surrogate to cut PDE solver calls in inversion, but the gradient fidelity claim rests on thin evidence. the 2 major comments →

arxiv 2606.26592 v1 pith:C5BQKQ27 submitted 2026-06-25 cs.CE cs.LG

Latent Diffusion Posterior Sampling with Surrogate Likelihood Guidance for PDE Inverse Problems

Yuanzhe Wang , Alexandre M. Tartakovsky This is my paper
classification cs.CE cs.LG
keywords latent diffusionposterior samplingPDE inverse problemsvariational autoencoderneural surrogateDarcy flowBayesian inversionparameter estimation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper develops L-DPS to sample from the posterior over high-dimensional parameter fields in PDE-governed inverse problems. A variational autoencoder compresses the parameters into latent space, an unconditional diffusion model supplies the prior score there, and diffusion posterior sampling is guided by likelihood gradients obtained from the decoder composed with a differentiable neural surrogate. This construction is demonstrated on recovering a permeability field from sparse noisy pressure measurements in Darcy flow, where the resulting samples are accurate and the inference cost drops relative to full-space diffusion posterior sampling.

Core claim

L-DPS performs diffusion posterior sampling in the latent space of a variational autoencoder, with the likelihood gradient supplied by the composition of the VAE decoder and a neural surrogate model of the forward PDE; the resulting posterior samples are obtained without repeated calls to the original numerical solver.

What carries the argument

The decoder-surrogate composition that supplies the likelihood gradient to latent diffusion posterior sampling.

Load-bearing premise

The neural surrogate must supply gradients accurate enough that the guided diffusion samples remain close to those obtained with the true PDE likelihood.

What would settle it

Replacing the surrogate with the true PDE solver inside the same sampling loop and observing that the posterior mean or covariance changes by more than sampling noise would show the surrogate guidance is not faithful.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • L-DPS produces accurate and robust reconstructions of the unknown parameter field from sparse noisy observations.
  • Inference cost drops relative to full-space diffusion posterior sampling because the expensive PDE solver is replaced by the surrogate after training.
  • L-DPS outperforms amortized baselines such as conditional latent diffusion and inverse FNO in sparse and noisy observation regimes.
  • The method supports mixed-prior generalization by changing the latent diffusion model while keeping the surrogate fixed.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same latent-space guidance could be applied to other inverse problems whose forward maps admit cheap differentiable emulators.
  • One could examine whether the latent dimension trades off reconstruction fidelity against sampling speed in a predictable way.
  • Extending the surrogate to output uncertainty estimates might allow the guidance term to reflect forward-model error explicitly.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 1 minor

Summary. The manuscript proposes latent diffusion posterior sampling (L-DPS) for high-dimensional PDE inverse problems. It combines a VAE to encode parameter fields into latent space, an unconditional latent diffusion model to learn the prior score, diffusion posterior sampling to incorporate likelihood guidance, and a differentiable neural surrogate so that the likelihood gradient is obtained via the decoder-surrogate composition rather than repeated full PDE solves. The method is demonstrated on an inverse Darcy-flow problem with sparse noisy pressure observations; the abstract claims that L-DPS yields accurate robust reconstructions, lowers inference cost relative to full-space DPS, outperforms amortized baselines (conditional latent diffusion, inverse FNO), and includes studies of mixed-prior generalization and sensitivity to surrogate forward-model error.

Significance. If the surrogate-gradient fidelity claim holds, the work supplies a practical route to Bayesian inversion when the prior is only available implicitly through samples and the forward map is expensive, by moving both the diffusion process and the likelihood correction into a learned latent space. The combination of VAE + latent DPS + surrogate is a concrete engineering contribution that could be reused in other PDE-constrained settings once the gradient-accuracy issue is closed.

major comments (2)
  1. [Abstract / likelihood-gradient paragraph] Abstract and the method section describing the likelihood term: the central claim that L-DPS produces samples from (an approximation to) the true PDE-constrained posterior rests on the assertion that the surrogate supplies gradients whose error does not materially distort the DPS score correction. The manuscript states that sensitivity to surrogate forward-model error is studied, yet supplies no quantitative bound on gradient discrepancy, no direct comparison of surrogate versus true-PDE gradients evaluated along the sampling trajectory, and no ablation that replaces the surrogate gradient with the exact one. This verification is load-bearing for the reported gains in sparse/noisy regimes.
  2. [Experiments / Darcy results] Experiments section (Darcy-flow results): without the controls listed above, it is impossible to determine whether the reported outperformance versus conditional latent diffusion and inverse FNO reflects faithful posterior approximation or systematic bias introduced by the surrogate. A single quantitative table or figure showing, e.g., the L2 norm of the gradient difference ||∇_θ log p(y|θ) – ∇_θ log p̂(y|θ)|| along sampled trajectories would directly address the issue.
minor comments (1)
  1. Notation: the manuscript would benefit from an explicit symbol (e.g., F̂) that distinguishes the neural surrogate from the true PDE operator F throughout the equations and algorithm box.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments highlighting the importance of validating the surrogate gradient approximation. We address each major comment below and agree that additional quantitative verification will strengthen the manuscript.

read point-by-point responses

  1. Referee: [Abstract / likelihood-gradient paragraph] Abstract and the method section describing the likelihood term: the central claim that L-DPS produces samples from (an approximation to) the true PDE-constrained posterior rests on the assertion that the surrogate supplies gradients whose error does not materially distort the DPS score correction. The manuscript states that sensitivity to surrogate forward-model error is studied, yet supplies no quantitative bound on gradient discrepancy, no direct comparison of surrogate versus true-PDE gradients evaluated along the sampling trajectory, and no ablation that replaces the surrogate gradient with the exact one. This verification is load-bearing for the reported gains in sparse/noisy regimes.

    Authors: We agree that a direct quantitative assessment of gradient discrepancy is important for substantiating the claims. While the manuscript reports a sensitivity study on inversion accuracy with respect to surrogate forward-model error, it does not include explicit L2-norm comparisons of surrogate versus true gradients along trajectories or an ablation with exact PDE gradients. We will add this verification in the revision: a new figure showing ||∇_θ log p(y|θ) – ∇_θ log p̂(y|θ)|| evaluated at multiple points along sampled trajectories, plus a limited ablation (on a subset of trajectories) replacing the surrogate gradient with the exact one where computationally feasible. revision: yes

  2. Referee: [Experiments / Darcy results] Experiments section (Darcy-flow results): without the controls listed above, it is impossible to determine whether the reported outperformance versus conditional latent diffusion and inverse FNO reflects faithful posterior approximation or systematic bias introduced by the surrogate. A single quantitative table or figure showing, e.g., the L2 norm of the gradient difference ||∇_θ log p(y|θ) – ∇_θ log p̂(y|θ)|| along sampled trajectories would directly address the issue.

    Authors: We concur that the suggested quantitative control is needed to isolate the effect of the surrogate. The existing sensitivity study addresses overall inversion accuracy but not trajectory-level gradient fidelity. We will incorporate the requested figure (L2 gradient discrepancy along trajectories) and reference it in the Darcy results section to allow readers to assess whether the reported gains are attributable to faithful posterior sampling or surrogate bias. revision: yes

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The paper constructs L-DPS as an explicit composition of VAE dimensionality reduction, unconditional latent diffusion for the prior score, DPS guidance, and a decoder-surrogate composition for the likelihood gradient. None of these steps reduce the target posterior, the reported accuracy metrics, or the cost savings to quantities defined by the method's own fitted parameters or by self-referential equations. The Darcy-flow evaluation, baseline comparisons (conditional latent diffusion, inverse FNO, KLE-MAP), mixed-prior tests, and explicit sensitivity study to surrogate error are presented as independent empirical checks rather than tautological predictions. No self-citation load-bearing steps, uniqueness theorems, or ansatz smuggling appear in the abstract or method outline.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract supplies insufficient detail to enumerate free parameters, axioms, or invented entities; the surrogate model and VAE latent space are treated as standard components without stated assumptions or new postulated objects.

reviewed 2026-06-26 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Latent Diffusion Posterior Sampling with Surrogate Likelihood Guidance for PDE Inverse Problems." pith.science (2026). https://pith.science/paper/C5BQKQ27

@misc{pith2026260626592,
  author       = {Pith},
  title        = {Pith review of: Latent Diffusion Posterior Sampling with Surrogate Likelihood Guidance for PDE Inverse Problems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/C5BQKQ27}},
  note         = {Machine review of arXiv:2606.26592}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

We propose latent-space diffusion posterior sampling (L-DPS), an approximate Bayesian framework for high-dimensional inverse problems governed by partial differential equations (PDEs). The method addresses three challenges in PDE-constrained inversion: implicit sample-based priors without tractable densities, high-dimensional spatially distributed parameters, and the high cost of repeated forward-model evaluations during posterior sampling. L-DPS combines a variational autoencoder, an unconditional latent diffusion model, diffusion posterior sampling, and a differentiable neural surrogate. The VAE maps the parameter field to a lower-dimensional latent space, the diffusion model learns an implicit prior score in this latent space, and DPS combines this learned prior with likelihood-based guidance. The likelihood gradient is evaluated through the decoder-surrogate composition, avoiding repeated calls to the full numerical PDE solver. We evaluate the method on an inverse Darcy flow problem with an unknown spatially distributed permeability field inferred from sparse and noisy pressure observations. L-DPS produces accurate and robust inverse solutions, reduces inference cost relative to full-space DPS, and outperforms amortized inverse baselines such as conditional latent diffusion and inverse FNO in sparse and noisy regimes. We further compare L-DPS with a KLE-MAP baseline and study mixed-prior generalization and the sensitivity of inversion accuracy to surrogate forward-model error.

Figures

Figures reproduced from arXiv: 2606.26592 by Alexandre M. Tartakovsky, Yuanzhe Wang.

Figure 1
Figure 1. Figure 1: Overview of the L-DPS framework. The VAE is trained to reconstruct the parameter field [PITH_FULL_IMAGE:figures/full_fig_p011_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Comparison of the reference parameter field ( [PITH_FULL_IMAGE:figures/full_fig_p014_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: summarizes the quantitative performance by plotting the mean relative ℓ2 error in the estimated y field as a function of the number of observations Nobs for four noise levels, γ ∈ {0, 0.01, 0.05, 0.1}. Across most observation and noise settings, L-DPS+FNO matches or outperforms the benchmark methods. The advantage is especially pronounced in noisy regimes, where enforcing data consistency without a suffici… view at source ↗
Figure 4
Figure 4. Figure 4: VAE reconstruction quality for three priors. Rows: single-scale Gaussian, multi-scale Gaussian, and binary [PITH_FULL_IMAGE:figures/full_fig_p017_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Distribution of the VAE reconstruction relative [PITH_FULL_IMAGE:figures/full_fig_p017_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Errors in the inverse L-DPS solutions versus errors in the FNO, ViT, and DeepONet (forward) surrogate [PITH_FULL_IMAGE:figures/full_fig_p018_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Inverse relative ℓ2 errors versus the number of observations Nobs for six guidance strategies across four noise levels γ ∈ {0.01, 0.02, 0.05, 0.1}. All methods share the same L-DPS+FNO pipeline and differ only in the guidance rule. 5.9 Comparison with the KLE-MAP method We compare L-DPS+FNO against the KLE-MAP baseline for the three different y priors described in Section 5.7. For Gaussian priors, KLE-MAP … view at source ↗
Figure 8
Figure 8. Figure 8: Inverse reconstructions at Nobs = 128, γ = 0.01 for three priors. Rows from top to bottom: Single-scale Gaussian, multi-scale Gaussian, Non-Gaussian binary. Columns from left to right: reference yref , L-DPS+FNO estimate yˆL−DP S, and point errors yˆL−DP S −yref . Solutions are shown for the test cases corresponding to the median ℓ2 errors. 22 [PITH_FULL_IMAGE:figures/full_fig_p022_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Inverse reconstructions at Nobs = 128, γ = 0.01 from a single L-DPS+FNO model trained on a mixed dataset. Rows correspond to the three source priors; the model is not told which prior a test sample came from. 23 [PITH_FULL_IMAGE:figures/full_fig_p023_9.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

13 extracted references · 5 canonical work pages · 3 internal anchors

  1. [1]

    Score-based generative modeling through stochastic differential equations.ICLR, 2021

    Yang Song et al. Score-based generative modeling through stochastic differential equations.ICLR, 2021

    2021

  2. [2]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022

    2022

  3. [3]

    Diffusion posterior sampling for general noisy inverse problems.ICLR, 2023

    Hyungjin Chung et al. Diffusion posterior sampling for general noisy inverse problems.ICLR, 2023

    2023

  4. [4]

    Surgin: Surrogate-guided generative inversion for subsurface multiphase flow with quantified uncertainty.arXiv preprint arXiv:2509.13189, 2025

    Zhao Feng, Bicheng Yan, Luanxiao Zhao, Xianda Shen, Renyu Zhao, Wenhao Wang, and Fengshou Zhang. Surgin: Surrogate-guided generative inversion for subsurface multiphase flow with quantified uncertainty.arXiv preprint arXiv:2509.13189, 2025

    work page arXiv 2025

  5. [5]

    Denoising diffusion probabilistic models.NeurIPS, 2020

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models.NeurIPS, 2020

    2020

  6. [6]

    Diffusion Posterior Sampling for General Noisy Inverse Problems

    Hyungjin Chung, Jeongsol Kim, Michael T Mccann, Marc L Klasky, and Jong Chul Ye. Diffusion posterior sampling for general noisy inverse problems.arXiv preprint arXiv:2209.14687, 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  7. [7]

    Denoising diffusion restoration models.Advances in neural information processing systems, 35:23593–23606, 2022

    Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming Song. Denoising diffusion restoration models.Advances in neural information processing systems, 35:23593–23606, 2022

    2022

  8. [8]

    Taming transformers for high-resolution image synthesis

    Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873–12883, 2021

    2021

  9. [9]

    Zero-shot text-to-image generation

    Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea V oss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. InInternational conference on machine learning, pages 8821–8831. Pmlr, 2021

    2021

  10. [10]

    Vae-dnn: Energy-efficient trainable-by-parts surrogate model for parametric partial differential equations.arXiv preprint arXiv:2508.03839, 2025

    Yifei Zong and Alexandre M Tartakovsky. Vae-dnn: Energy-efficient trainable-by-parts surrogate model for parametric partial differential equations.arXiv preprint arXiv:2508.03839, 2025

    work page arXiv 2025

  11. [11]

    Denoising Diffusion Implicit Models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models.arXiv preprint arXiv:2010.02502, 2020

    work page internal anchor Pith review Pith/arXiv arXiv 2010

  12. [12]

    Jacobian-Aware Posterior Sampling for Inverse Problems

    Liav Hen, Tom Tirer, Raja Giryes, and Shady Abu-Hussein. Robust posterior diffusion-based sampling via adaptive guidance scale.arXiv preprint arXiv:2511.18471, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  13. [13]

    Diffusion state-guided projected gradient for inverse problems

    Rayhan Zirvi, Bahareh Tolooshams, et al. Diffusion state-guided projected gradient for inverse problems. In International Conference on Learning Representations, volume 2025, pages 31217–31242, 2025. 25 APREPRINT- JUNE26, 2026 A Hyperparameters Table 6 lists the hyperparameters of the V AE, Table 7 those of the latent UNet diffusion prior, and Table 8 tho...

    2025

This paper was first reviewed by grok-4.3 on June 26, 2026.