{"id":"5323daf5-5d80-4de0-857f-de7fe8408320","arxiv_id":"2509.04860","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Latent-space plug-and-play diffusion posterior sampling yields state-of-the-art synthetic EM brain image reconstructions without paired measurement-label training data.","lead":"This paper adapts plug-and-play diffusion posterior sampling to a learned latent space and applies it to electromagnetic brain imaging. It reports higher reconstruction accuracy and structural similarity than six baselines on synthetic stroke-imaging tests.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The inherited posterior-sampling guarantee is asserted, not verified, for the nonlinear latent-space ISP; the reported SOTA numbers are empirical and could survive, but the theoretical foundation of the MMSE/uncertainty claim is unsupported.","rationale":"The paper is a coherent and well-tested extension of DPnP to a latent space for EM inverse scattering, with strong empirical results on synthetic benchmarks and a careful comparison against seven methods. The reader's weakest assumption correctly identifies the unsupported transfer of the convergence guarantee to the latent-space setting. I agree with that core concern and broaden it: the nonlinear EM forward operator itself, not only the latent-space decoder, is never checked against the assumptions of [47] and [48]. I also flag a concrete sign inconsistency between the SDE (23) and its discretization (24), which indicates the theoretical derivation was not carefully validated. None of this refutes the reported Table I numbers, because the empirical claim can in principle stand even if the samples are not exactly from the posterior. However, the paper explicitly claims a posterior-sampling method with theoretical guarantees, and the MMSE and uncertainty products depend on that claim. Therefore the appropriate outcome is a conditional acceptance requiring either a proof of the latent-space nonlinear extension, a clear statement of the assumptions that are inherited, or a revision of the theoretical framing. Since the reader already reached CONDITIONAL, my assessment does not change the verdict.","tokens_in":26445,"tokens_out":14175,"duration_ms":142523,"concrete_test":"Check the theorem statements in [47] and [48] for their assumptions (linear or Lipschitz forward operator, bounded domain, bounded score error, exact proximal sampling). Determine whether those assumptions hold for F∘G in Eq. (5) with the VAE decoder and for the diffusion-model approximation of q1 in latent space. If any condition fails, the claimed extension to the nonlinear EM ISP is unproven; the authors should either provide a new convergence theorem for this setting or explicitly soften the posterior-sampling and MMSE claims. As an empirical adjunct, rerun the ATLAS experiment with a Born-linearized forward model in place of (5); if the reconstructions and uncertainty maps change substantially, the nonlinearity is a live part of the sampling dynamics and the missing guarantee is material.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III-C inherits the alternating-proximal convergence guarantee from Bouman and Buzzard [47] and Xu and Chi [48] without establishing that its conditions hold here. The likelihood (15) uses the EM forward map F of (5), which is nonlinear through the matrix inverse (I - G_D \\chi)^{-1}, composed with a VAE decoder G. The cited guarantees, at least as stated for DPnP in [48], are established for specific (typically linear or mildly nonlinear) measurement models and require regularity conditions; the paper supplies no verification that F∘G satisfies them. The move to a learned latent space also changes the prior step from a known proximal distribution to a diffusion-model approximation with unquantified score error, and no error bound is given for this adaptation. In the implementation, the guarantees are further weakened by a finite schedule (η_k decreasing only to 0.03 for ATLAS, N_k=20, M=5) and by the adaptive gradient rescaling α_n in (37), so the samples are not demonstrated to be from p(z|d_obs); the MMSE estimate (12) and the uncertainty maps in Fig. 8 inherit this gap. There is also an internal inconsistency: SDE (23) has a negative sign on ∇L, while the exponential integrator (24) corresponds to the opposite sign, so the written derivation of the likelihood step does not match the implemented update. These issues do not directly disprove the reported quantitative results, but they undermine the paper's claim of posterior sampling with theoretical guarantees and reliability/uncertainty quantification.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces L-DPnP, a latent-space plug-and-play posterior sampling algorithm for electromagnetic inverse scattering. A variational autoencoder is trained on unlabeled pairs of permittivity and conductivity maps, a score-based latent diffusion model is trained on the encoded latents, and at inference each posterior sample is produced by alternating a likelihood step (a Langevin/exponential-integrator update through the composed forward model F∘G) with a diffusion-based prior step; M decoded samples are averaged to approximate the MMSE estimate. The method is evaluated on MNIST, Fashion-MNIST, and ATLAS brain phantoms, where it reports the best reconstruction RMSE and SSIM among seven methods while keeping measurement RMSE near the noise level, and uncertainty maps are shown.","tokens_in":26752,"tokens_out":7403,"duration_ms":62474,"significance":"If the results hold, the paper describes a practically useful framework: it is, to my knowledge, the first application of latent-space diffusion posterior sampling to nonlinear electromagnetic inverse scattering, it avoids paired measurement-label training for the prior, and it is modular with respect to measurement geometry. The experimental comparison is solid in that all seven methods use the same forward model and the same test set, 200 cases per dataset, and the quantitative gains on ATLAS are substantial (reconstruction RMSE 0.075 versus 0.086 for the second-best method). The paper also provides detailed pseudocode and architecture descriptions. The main weakness is that the theoretical guarantee underlying the 'posterior sampling' framing is asserted rather than verified, and the implemented likelihood step contains a sign inconsistency and a heuristic adaptive rescaling; these issues do not disprove the empirical numbers, but they do undermine the theoretical claims as currently stated.","major_comments":[{"comment":"The SDE written in (23) has dzτ = −∇L dτ + (1/η_k^2)(zτ−ż_k)dτ + √2 dwτ, but the exponential integrator (24) with r=e^{−γ/η^2} is the exact discretization of dzτ = +∇L dτ − (1/η_k^2)(zτ−ż_k)dτ + √2 dwτ. The signs of both the gradient term and the linear drift are opposite. Since Algorithm 2 and the implementation in Appendix B use (24), the derivation of the likelihood step as sampling q0 in (17) is internally inconsistent. The correct SDE for q0 is the one with the plus sign on ∇L and the minus sign on the linear drift, so the text should be corrected; as written, a reader cannot reproduce the algorithm from the equations.","section":"Section IV-A, Eqs. (23)-(24)"},{"comment":"The convergence guarantee is imported from [47] and [48] without checking its hypotheses in the present setting. The likelihood here is the nonlinear composed map F∘G, where F includes the matrix inverse (I−G_D χ)^{−1} of (4), and the prior step is implemented by a learned score model rather than by the exact proximal distribution q1. Neither the regularity conditions of [48] (for example, Lipschitz or bounded-gradient assumptions, or linear-Gaussian structure) nor an error bound for the score approximation is stated. Moreover, the finite annealing schedule (η_k decreases only to 0.03 for ATLAS, with N_k=20 and M=5) is far from the asymptotic regime η→0, N_k→∞ used in the cited theory. Therefore the samples are not demonstrated to come from p(z|d_obs), and the MMSE estimate (12) and the uncertainty maps in Fig. 8 inherit this gap. This should be fixed either by proving the relevant conditions or by explicitly presenting the method as a heuristic sampler whose posterior accuracy is validated empirically.","section":"Section III-C and Section VI-A"},{"comment":"The adaptive coefficient α_n in (37) multiplies the gradient term by a state-dependent factor proportional to ||z[n]||^2 / (||∇L(z[n])||^2 + 0.001). This rescaling is not the gradient of a fixed potential, and it changes the invariant measure of the likelihood-step chain; the choice of α_0 and the per-dataset η_k annealing schedules is hand-tuned. Consequently, even after the sign in (23) is corrected, the implemented likelihood step does not sample the proximal distribution q0 in (17). The paper should either justify this update as a valid MCMC kernel (for example, a preconditioned Langevin update with a known stationary distribution) or present it as a heuristic acceleration whose effect on the posterior is empirically assessed.","section":"Appendix B7, Eqs. (36)-(37)"},{"comment":"Because the posterior-sampling guarantee is not established, and because the uncertainty maps are shown only for a few examples, the claim of 'reliable reconstruction' through MMSE and uncertainty quantification is under-supported. I would like to see a posterior predictive check, such as the coverage of credible intervals over the 200 test cases or a comparison between sample variance and actual reconstruction error, to demonstrate that the sample spread has calibration meaning despite the heuristic adaptations. Without such validation, Fig. 8 is an illustration rather than evidence of calibrated uncertainty.","section":"Section V-C and Fig. 8"}],"minor_comments":[{"comment":"The second term on the right-hand side reads 'r x[n]' but should be 'r z[n]' to match the latent-space notation used elsewhere.","section":"Appendix B7, Eq. (36)"},{"comment":"The symbol n is used both as the discrete-time index and as the Gaussian noise vector; renaming one of them (for example, using ε for the noise realizations) would remove ambiguity.","section":"Algorithm 2, lines 11 and 20"},{"comment":"The likelihood variance σ^2 in (15)-(16) is never specified in the experimental section; please state how it is set relative to the noise level in (31), since this parameter directly controls the strength of the likelihood step.","section":"Section II-A and Eq. (15)"},{"comment":"The paper reports that L-DPnP keeps measurement RMSE 'near the noise level'; because the noise in (31) is defined separately for the real and imaginary parts, a brief explanation of how the ATLAS value 0.192 compares with the 20% noise level would help interpret the table.","section":"Table I and Section V-C"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is in scope for the journal, and the empirical study is substantial. The main risk is overclaiming the theoretical foundation: the sign inconsistency in Eq. (23), the unverified import of the convergence guarantee for the nonlinear latent setting, and the state-dependent rescaling in Eq. (37) need to be addressed before publication. The self-citations to [41] and [50] are appropriate because they supply the architecture and preliminary results on which the present work builds."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing to know: this is a serious applied ISP paper, not a theoretical one. The actual new bit is putting the plug-and-play diffusion posterior sampler into the latent space of an autoencoder for EM inverse scattering, with the argument that latent coordinates equalize the wildly varying sensitivity of the pixel-domain forward map. On MNIST, Fashion-MNIST, and ATLAS-derived brain slices, L-DPnP beats six baselines on reconstruction RMSE/SSIM while staying near the noise floor on measurement RMSE. The experiments look carefully run: 200 test cases per dataset, same forward model for all methods, and the comparison set includes learning-free, model-based deep learning, and PnP baselines. That is real work and the improvement is consistent.\n\nWhat is less solid is the theoretical wrapper. The convergence guarantee is inherited from Bouman/Buzzard and Xu/Chi without any demonstration that the relevant conditions hold for the nonlinear decoder forward map F∘G, or that the diffusion prior step has the required denoiser properties. The abstract and intro sell this as a method with theoretical guarantees; the body just cites them. I think the empirical result can stand without that guarantee, but the claim as written is overreach. There is also a sign inconsistency between the SDE (23) and the exponential integrator (24); from the algebra, (24) is the correct discretization and (23) has the wrong sign on ∇L. Quick fix, but it should be caught. On the practical side, the method leans on per-dataset annealing schedules and α_n scaling, there are no error bars on the reported metrics, and no code or data are released. With M=5 samples, the 'uncertainty' maps in Fig. 8 are illustrative at best, not calibrated UQ.\n\nNone of these problems kills the central claim—the reconstruction numbers are what they are on synthetic data—but they do mean the paper is not as clean as its framing suggests. It deserves a serious referee: a good reviewer will ask the authors to either prove or soften the convergence claim, fix the SDE sign, release code/data, and report variance across seeds. If those revisions land, it would be a useful contribution to the EM imaging and PnP literature. I would not desk-reject it.","headline":"Solid empirical paper on latent-space DPnP for EM inverse scattering; the theoretical guarantee is inherited rather than proven, and the UQ is thinner than claimed, but the reconstruction results deserve a serious referee.","tokens_in":27273,"tokens_out":3723,"would_cite":true,"duration_ms":34260,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A latent-space plug-and-play diffusion sampler achieves state-of-the-art quantitative electromagnetic brain imaging without paired training data.","keywords":["inverse scattering problems","brain imaging","posterior sampling","latent diffusion","plug-and-play","MMSE estimation","uncertainty quantification","electromagnetic imaging"],"falsifier":"Run L-DPnP on synthetic brains with known ground truth, and compare its posterior samples with an exact reference posterior computed by a slower Markov-chain sampler on a small problem where that is feasible. If the L-DPnP sample mean is biased relative to the reference, or if its 90% credible intervals cover the true map in well under 90% of test cases, then the latent-space adaptation of the proximal-sampling guarantee is not holding and the reported RMSE advantage would not reflect true posterior sampling.","tokens_in":26273,"feed_emoji":"🧠","tokens_out":10524,"duration_ms":85498,"temperature":0.7,"pith_summary":"This paper tries to establish that a latent-space plug-and-play diffusion sampler, called L-DPnP, can solve the nonlinear and ill-posed electromagnetic inverse scattering problem accurately enough for stroke brain imaging. Unlike methods that need paired measurement-label datasets, L-DPnP learns the prior distribution of permittivity and conductivity maps from unlabeled images, then enforces agreement with measured scattered fields through repeated likelihood and prior sampling steps. The paper argues that sampling in a learned latent space fixes the slow convergence that makes pixel-domain diffusion plug-and-play impractical for EM imaging, because latent coordinates have more balanced sensitivity to the measurements. If correct, the approach would give clinicians a non-ionizing, portable imaging modality that returns both a reconstruction and a per-pixel uncertainty map.","feed_headline":"Latent diffusion beats seven baselines on brain EM imaging","feed_subtitle":"Unlabeled brain scans teach the prior; physics keeps measurement fit, yielding the lowest reconstruction error.","key_machinery":"The load-bearing object is the learned latent space: a VAE encoder $E$ maps paired permittivity $\\epsilon_r$ and conductivity $\\sigma_e$ maps to a low-dimensional latent map $z$, and a decoder $G$ maps the latent back to physical maps. The Bayesian posterior is re-expressed as $p(z|d_{\\mathrm{obs}}) \\propto p(d_{\\mathrm{obs}}|z)\\,p(z)$, where the likelihood uses the composed nonlinear forward model $F(G(z))$ and $p(z)$ is captured by a score-based latent diffusion model. Sampling alternates a likelihood step, which draws from a proximal distribution $q_0(z|\\hat{z}^k;\\eta_k)$ via a Langevin SDE discretized with an exponential integrator, and a prior step, which draws from $q_1(z|\\hat{z}^{k+1/2};\\eta_k)$ by running the reverse-time diffusion SDE as a denoiser. The MMSE estimate is the average of $M$ decoded latent samples.","core_discovery":"L-DPnP replaces pixel-domain diffusion plug-and-play sampling with sampling in the latent space of a trained autoencoder, and claims that this change makes posterior sampling practical for nonlinear electromagnetic inverse scattering. On a 200-case ATLAS brain-imaging test the paper reports reconstruction RMSE of 0.075 and SSIM of 0.922, the best among the seven compared methods, while measurement RMSE of 0.192 stays close to the 20% noise level, indicating reconstructions that are both accurate and physically consistent. The same ranking pattern holds on MNIST and Fashion-MNIST-derived benchmarks. The paper attributes the improvement to a learned, material-aware latent representation that balances measurement sensitivities across latent coordinates, and to MMSE estimation over posterior samples rather than a single MAP reconstruction.","pith_inferences":["Beyond the paper, the latent-space sensitivity-balancing argument should transfer to other nonlinear inverse problems whose forward sensitivity varies across the domain, such as ultrasound tomography or electrical impedance tomography; the paper names these as future applications but does not test them.","Beyond the paper, a controlled ablation that fixes the diffusion prior and varies only the autoencoder's latent dimension could separate how much of the reported gain comes from balanced gradients versus how much comes from the expressiveness of the learned prior.","Beyond the paper, the claim that MMSE averaging avoids overconfident MAP solutions could be tested statistically on synthetic brains by comparing L-DPnP's posterior mean with single MAP draws across many cases, and by checking whether per-pixel variance tracks actual reconstruction error.","Beyond the paper, since the convergence guarantee in [47] and [48] is proven for the pixel-domain setting, a natural falsifiable check is to compare L-DPnP samples on a small EM problem against brute-force MCMC to test posterior calibration."],"forward_implications":["On the ATLAS brain dataset, L-DPnP reports the lowest reconstruction RMSE (0.075) and highest SSIM (0.922) among the seven compared methods, with measurement RMSE (0.192) near the 20% noise level.","Because the prior is learned from unlabeled property maps rather than from measurement-label pairs, the same trained prior can be reused when the antenna configuration or operating frequency changes.","The sampler outputs multiple posterior draws, so the reconstruction comes with a per-pixel standard deviation map that flags unreliable regions.","Working in the latent space balances the sensitivity of the unknown parameters to the measurements, cutting the number of alternating iterations from a practically infeasible value (about 200, ~6 hours) to 20 (~25 minutes) in the reported setup.","MMSE averaging over samples avoids the overconfident single-map solutions that MAP-based methods such as GMR can produce."],"supporting_citations":[{"why":"Supplies the alternating proximal-distribution Markov chain result that turns intractable posterior sampling into separate likelihood and prior sampling steps.","marker":"[47]"},{"why":"Defines diffusion plug-and-play posterior sampling with a convergence guarantee; its pixel-domain version (P-DPnP) is the main baseline L-DPnP builds on and is compared against.","marker":"[48]"},{"why":"Extends principled probabilistic imaging with diffusion plug-and-play priors, supporting the posterior-sampling interpretation used here.","marker":"[49]"},{"why":"Provides the VAE feature-space reparameterization for microwave brain imaging, the tissue electrical-property assignments for the ATLAS data, and the GMR baseline.","marker":"[41]"},{"why":"Supplies the latent diffusion modeling recipe and the argument that lower-dimensional latent sampling is more efficient than pixel-domain diffusion.","marker":"[67]"},{"why":"Provides the Green's function formulation of the forward scattering model and the back-projection method used for baseline comparisons.","marker":"[12]"},{"why":"Establishes the limited degrees of freedom of scattered fields, which is the paper's justification for the ill-posedness of the inverse problem.","marker":"[13]"},{"why":"Introduces the plug-and-play priors concept that lets a learned denoiser be inserted into a physics-based reconstruction loop.","marker":"[37]"},{"why":"Introduces generative model reparameterization for inverse problems, the basis of the GMR baseline that L-DPnP is compared against.","marker":"[38]"},{"why":"Provides the score-based SDE framework used to train the latent diffusion score model and to implement the reverse-time prior sampler.","marker":"[54]"}],"fun_headline_variants":["Latent diffusion cuts brain EM imaging error","No paired data needed: latent diffusion for brain EM","Posterior sampling improves EM brain scans","Plug-and-play latent diffusion wins on brain EM","Latent diffusion achieves best brain EM reconstruction"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything rests on the assumption that the alternating proximal sampler still converges to the true posterior $p(z|d_{\\mathrm{obs}})$ when the prior step is a latent diffusion model inside a nonlinear autoencoder decoder; the paper provides no error bound for that adaptation.","fun_headline_variants_meta":{"raw":{"variants":["Latent diffusion cuts brain EM imaging error","No paired data needed: latent diffusion for brain EM","Posterior sampling improves EM brain scans","Plug-and-play latent diffusion wins on brain EM","Latent diffusion achieves best brain EM reconstruction"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000169,"raw_usage":{"total_tokens":1278,"prompt_tokens":975,"completion_tokens":303,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":591,"completion_tokens_details":{"reasoning_tokens":235}},"tokens_in":591,"tokens_out":303,"duration_ms":3063,"temperature":1.0,"reasoning_tokens":235,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T16:25:28.274017+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run L-DPnP on synthetic brains with known ground truth, and compare its posterior samples with an exact reference posterior computed by a slower Markov-chain sampler on a small problem where that is feasible. If the L-DPnP sample mean is biased relative to the reference, or if its 90% credible intervals cover the true map in well under 90% of test cases, then the latent-space adaptation of the proximal-sampling guarantee is not holding and the reported RMSE advantage would not reflect true posterior sampling.","supporting_citations":[{"cited_title":"Generative plug and play: Posterior sampling for inverse problems,","cited_arxiv_id":null,"evidence_quote":"Supplies the alternating proximal-distribution Markov chain result that turns intractable posterior sampling into separate likelihood and prior sampling steps."},{"cited_title":"Provably robust score-based diffusion posterior sampling for plug-and-play image reconstruction,","cited_arxiv_id":null,"evidence_quote":"Defines diffusion plug-and-play posterior sampling with a convergence guarantee; its pixel-domain version (P-DPnP) is the main baseline L-DPnP builds on and is compared against."},{"cited_title":"Prin- cipled probabilistic imaging using diffusion models as plug-and-play priors,","cited_arxiv_id":null,"evidence_quote":"Extends principled probabilistic imaging with diffusion plug-and-play priors, supporting the posterior-sampling interpretation used here."},{"cited_title":"Three dimensional microwave data inversion in feature space for stroke imaging,","cited_arxiv_id":null,"evidence_quote":"Provides the VAE feature-space reparameterization for microwave brain imaging, the tissue electrical-property assignments for the ATLAS data, and the GMR baseline."},{"cited_title":"High- resolution image synthesis with latent diffusion models,","cited_arxiv_id":null,"evidence_quote":"Supplies the latent diffusion modeling recipe and the argument that lower-dimensional latent sampling is more efficient than pixel-domain diffusion."},{"cited_title":"Plug-and-play priors for model based reconstruction,","cited_arxiv_id":null,"evidence_quote":"Introduces the plug-and-play priors concept that lets a learned denoiser be inserted into a physics-based reconstruction loop."},{"cited_title":"Compressed sensing using generative models,","cited_arxiv_id":null,"evidence_quote":"Introduces generative model reparameterization for inverse problems, the basis of the GMR baseline that L-DPnP is compared against."}],"review_version":2}