{"id":"41e8e385-2cf9-4512-a22d-32278b279d5b","arxiv_id":"2504.12154","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":2.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A review of deep generative model techniques for Bayesian inverse problems in high-rate sensor data, covering structured noise, companding, model-based scores, and acceleration methods.","lead":"This paper reviews how deep generative models, mainly diffusion models, can be adapted for Bayesian inference on fast and noisy sensor data from automotive radar and medical imaging. It organizes known methods for handling model mismatch and real-time constraints, so engineers and researchers can see which approaches exist and where gaps remain.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Joint posterior sampling assumes p(n|x)≈p(n); in the ultrasound multipath example n depends on x, so the learned marginal noise prior is misspecified and posterior samples can be biased.","rationale":"The reader's weakest-assumption identification is correct: the independence assumption is explicitly introduced in Section 2(a), explicitly acknowledged to be false for multipath ultrasound, and then silently used to derive the factorized joint posterior in Eq (3.1) and the coupled diffusion updates in Eqs (3.3)-(3.4). This is the deepest load-bearing point because the structured-noise framework is the paper's primary proposed remedy for model mismatch. The reader also notes a concrete issue in Eq (3.8), where companding appears to mix RF and companded domains; I regard that as a real but likely correctable notational/domain error, whereas the independence assumption affects the whole methodology and cannot be fixed by rewriting one equation. The paper's own text contains the limitation, but it does not flag failure modes or give calibration evidence, so a conditional recommendation with a request for explicit discussion and validation is appropriate. Since the reader already gave CONDITIONAL, my stress test does not move the verdict.","tokens_in":16108,"tokens_out":6692,"duration_ms":71038,"concrete_test":"Run the Box 3.1 ultrasound dehazing pipeline on simulated RF data in which multipath haze is generated conditionally on the scene, e.g. n=H(x)+z with H a known multipath operator (delayed and attenuated copies of the transmitted pulse convolved with x) and z an independent component. Train sθ and sϕ on marginals as in [32], perform joint posterior sampling via Eqs (3.2)-(3.4), and measure posterior-mean RMSE and empirical coverage of 90% credible intervals for x as the fraction of haze energy coming from H(x) is increased from 0 to 1. If coverage drops well below 90% as the conditional component grows, the independence assumption is empirically load-bearing; if coverage stays near nominal, the marginal-prior approximation is benign in this setting.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The structured-noise framework in Section 3(a) is the paper's central answer to model mismatch, and it rests on a prior independence assumption stated in Section 2(a): after defining y=Ax+n+ε, the text says 'n is inherently a function of x, e.g. n(x)' but then sets p(n|x)≈p(n) and learns the marginal p(n). This assumption is what justifies the factorized joint posterior in Eq (3.1), p(x,n|y)∝p(y|x,n)p(x)p(n), and the coupled reverse-time SDE in Eqs (3.2)-(3.4). For the ultrasound multipath example in Box 3.1, this is not a harmless approximation: haze n is generated by the same transmitted pulse scattering off the tissue x, so p(n|x) is strongly coupled to x. Training separate score models sθ(xτ) and sϕ(nτ) on unpaired 'clean' and 'haze' samples only captures the marginal p(n); it cannot represent the conditional relationship. At inference the likelihood term must carry the full burden of enforcing consistency, and with a misspecified prior the DPS-style posterior sampler can produce biased point estimates and overconfident credible intervals. The paper flags the assumption but never discusses its failure modes or provides validation that posterior samples remain calibrated in exactly the regime (multipath) where it is invoked. Since this framework is also used for radar interference mitigation, the concern is load-bearing for the review's claim that these techniques address model mismatch.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript is a review of deep generative model (DGM) based Bayesian inference for high-rate sensor data, with applications to automotive radar and medical imaging. It introduces a forward model y = Ax + n + ε, reviews score-based diffusion posterior sampling via Tweedie's formula, and then discusses three classes of methods for model mismatch: modeling structured noise with separate DGMs, companding to handle high dynamic range, and model-based score functions. It also reviews acceleration strategies (temporal initialization, deep unfolding, knowledge distillation) and active compressed sensing. The central claim is that these techniques make DGM-based posterior sampling applicable to real-time sensing, illustrated with ultrasound multipath dehazing, radar interference mitigation, and accelerated MRI.","tokens_in":16364,"tokens_out":7105,"duration_ms":59307,"significance":"The paper's background derivation of posterior sampling is clean and the survey of active compressed sensing covers several recent external methods, which gives it value as a structured introduction to the area. If the reviewed techniques are correct, they would indeed extend DGM-based Bayesian inference to high-rate sensing. However, the technical credibility of the main model-mismatch framework is undermined by a factorized joint posterior that relies on an independence assumption the paper itself contradicts for the ultrasound example, and by an algebraic error in the companding likelihood (Eq. (3.8)). These issues are fixable, but they are load-bearing for the paper's central claims.","major_comments":[{"comment":"The structured-noise framework assumes p(n|x) ≈ p(n) and learns only the marginal p(n), which is what justifies the factorized joint posterior in Eq. (3.1) and the coupled reverse-time SDEs in Eqs. (3.3)-(3.4). However, Section 2(a) itself states that in ultrasound multipath scattering, 'n is inherently a function of x', and Box 3.1 models clutter and multipath components as n generated by the same transmitted pulse. In this regime the independence assumption is misspecified: unpaired training of separate score models s_θ(x_τ) and s_ϕ(n_τ) captures only marginals, and the likelihood term must carry the full burden of enforcing consistency. The manuscript does not analyze the resulting bias in posterior samples, nor does it provide any calibration or validation study for exactly the multipath regime where the assumption is invoked. Since this framework is also used for radar interference mitigation, this concern is load-bearing for the claim that the surveyed techniques address model mismatch.","section":"Section 2(a), Eq. (3.1), Box 3.1"},{"comment":"The second equality in Eq. (3.8), C(A x_RF,0|τ + n_RF,0|τ) = C(C^{-1}(A x_{0|τ}) + C^{-1}(n_{0|τ})), is not correct. The companding functions C and C^{-1} act elementwise on RF-domain vectors, while A x_{0|τ} is a measurement-domain vector; applying C^{-1} to A x_{0|τ} is not a defined operation in this setup, and A and C^{-1} do not commute (A C^{-1}(x) ≠ C^{-1}(A x) in general). Moreover, if y in Eq. (3.7) is the raw RF measurement, the likelihood residual should be y - (A C^{-1}(x_{0|τ}) + C^{-1}(n_{0|τ})), not y - C(A C^{-1}(x_{0|τ}) + C^{-1}(n_{0|τ})). This invalidates the data consistency term used to produce the ultrasound results in Fig. 2 unless the variables are redefined and the derivation is corrected.","section":"Section 3(b), Eq. (3.8)"}],"minor_comments":[{"comment":"The gradient on the right-hand side of Eq. (3.7) should be with respect to both x_τ and n_τ, not only x_τ; as written, the update for n_τ is undefined.","section":"Section 3(b), Eq. (3.7)"},{"comment":"The notation x_RF,0|τ and x_{0|τ} is introduced without clearly stating whether these are Tweedie estimates in the RF domain or in the companded domain; please define both explicitly before Eq. (3.8).","section":"Section 3(b)"},{"comment":"The text refers to 'principle components' of the posterior covariance; this should be 'principal components'.","section":"Section 4(b), AdaSense paragraph"},{"comment":"The phrase 'hundreds or thousands Gigabits per second' should be 'hundreds or thousands of gigabits per second'.","section":"Box 3.1"},{"comment":"Knowledge distillation is described as a promising acceleration method, but no sensing-specific example or reference to high-rate sensing applications is given; the review should either provide one or explicitly state that this remains future work.","section":"Section 4(a)(iii)"}],"recommendation":"major_revision","confidential_remarks":"The paper's main worked examples are drawn almost exclusively from the authors' own prior publications ([31], [32], [33], [47], [68]) and no independent code or data artifacts are provided; for a review this is acceptable, but the editor may wish to confirm with the authors that the cited results are published or in press. The independence assumption and Eq. (3.8) are the two points I would most want addressed; the second appears to be a genuine technical error that the authors should correct carefully rather than dismiss as notation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear [Colleague],\n\nThis is a solid, readable review of the authors' own line of work on deep generative models for Bayesian inference in high-rate sensing, focused on automotive radar and medical ultrasound. The useful part is the organization: separating model mismatch from real-time constraints and showing how structured-noise modeling, companding, temporal initialization, deep unfolding, and active compressed sensing fit together. The boxes with data rates and forward models are genuinely helpful for engineers. The Tweedie/DPS background in Section 2 is clean and accurate.\n\nThe main soft spot is Eq (3.8). As written, it applies the companding function C to the sum of expanded components inside the data-consistency term. Since y is raw RF data and C maps RF to companded domain, the expression mixes domains. The correct gradient should be something like ∇ || C(y) - (A x + n) ||² if the generative model lives in the companded domain, or ∇ || y - (A C^{-1}(x) + C^{-1}(n)) ||² if the forward model is in RF; either way, the noise statistics need to be handled. This needs a fix or a clarifying remark.\n\nThe deeper issue is the independence assumption p(n|x) ≈ p(n), stated in Section 2(a) and used for the joint posterior in Eq (3.1). The paper is honest enough to say n can be a function of x, but it never explores the consequences. For ultrasound multipath, the haze comes from the same transmit pulse scattered off tissue, so the coupling is physical. Training a marginal score model for n will miss that. The result could be biased posterior means and overconfident intervals, exactly where the method is supposed to help. I wouldn't call this paper-wrecking, because the assumption is explicit and the empirical results may be fine, but a review that claims to address model mismatch should include a critical paragraph on when independence is defensible and what to do when it isn't.\n\nAlso note the heavy self-citation: most worked examples are the authors' prior papers, with no independent artifacts. For a review this is not disqualifying, but balancing with external methods (they do cite Sanchez, van de Camp, Elata, Kawar) and a sentence about limitations would help.\n\nWho is this for? Engineers and applied researchers entering DGM posterior sampling; they'll get a good map. I recommend sending to peer review, with a request to fix Eq (3.8) and add the independence caveat. It would probably be accepted after that.","headline":"A competent but heavily self-referential review of DGM-based Bayesian inference for radar and ultrasound; the synthesis is useful, but Eq (3.8) mixes domains and the core independence assumption needs a critical paragraph.","tokens_in":16942,"tokens_out":4657,"would_cite":true,"duration_ms":40901,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":null,"created_at":"2026-08-16T12:36:52.626286+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":null,"supporting_citations":[],"review_version":1}