{"id":"cb09c9bd-3d40-416e-ba5b-076acda71fa6","arxiv_id":"2607.14597","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Adaptive SC-VAMP with local tilted-EM updates self-tunes hyperparameters and reaches near-oracle compressed-sensing performance from strongly mismatched starts.","lead":"A message-passing algorithm for recovering hidden signals gets a new built-in tuning step: each of its components re-estimates its own assumptions while the algorithm runs. In compressed-sensing tests it reaches almost the accuracy of a system that already knew the true settings, even when starting far off.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Proposition 1 only proves a fixed point, not convergence; the 'self-tuning' claim hinges on unanalyzed stability of the population recursion.","rationale":"Proposition 1 is internally sound: the prior-module proof is a clean KL argument, and the likelihood-module proof reduces to standard identifiability under the cavity law. However, the paper stops at fixed-point existence. In EM-style self-tuning, a fixed point can be unstable or have a small basin; without a spectral-radius or contraction analysis, the title's 'self-tuning' is not supported. The numerical section gives only two instantiations, no code, no existing-adaptive-AMP baselines, and no convergence analysis. The one-bit experiment's SNR cutoff at 10 dB acknowledges an identifiability boundary, which itself shows that the theoretical assumptions are not universally satisfied. A concrete Jacobian/state-evolution test would settle whether the fixed point is stable in the claimed regimes. Since the reader's CONDITIONAL verdict already captures the need for additional analysis, no change in verdict is needed.","tokens_in":7328,"tokens_out":20893,"duration_ms":222410,"concrete_test":"Compute the Jacobian J = D_{(θ,v)}(M, F) at (θ0, v⋆) for the linear and one-bit BG instantiations from the state-evolution equations in Section II-D. If the spectral radius of J exceeds 1, the matched fixed point is unstable and the algorithm cannot self-tune from nearby mismatched initializations. If it is <1, run the deterministic population recursion (11) from 1000 random initializations drawn from the Section III ranges and measure the fraction that converge to a neighborhood of (θ0, v⋆). This directly settles whether the missing convergence analysis is a practical failure or merely a theoretical gap.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central theoretical result, Proposition 1, establishes that (θ0, v⋆) is a fixed point of the population map (θ, v) ↦ (M(θ, v), F(v; θ)). It does not establish that the adaptive recursion converges to this fixed point from mismatched initialization. The paper's claim that the method 'self-tunes' is therefore supported only by the two numerical experiments in Section III, which use a limited set of initializations, omit baselines (EM-GAMP, EM-VAMP, EM-GVAMP), and provide no code. More importantly, the likelihood-module proof requires the 'local compatibility condition' (Section II-D): that the cavity-weighted conditional KL identifies θ0. This is assumed, not derived. For the linear Gaussian likelihood the condition does hold, but for the one-bit probit factor the paper itself shows an identifiability boundary (SNR≈10 dB) where the sign map is nearly scale-invariant and σ_x² is not reliably identifiable; at that boundary the assumptions of Proposition 1 are not met. The fixed-point theorem is correct as a conditional statement, but the gap between a fixed point and self-tuning convergence is the load-bearing weak point: an unstable fixed point would invalidate the titular claim even though Proposition 1 remains true.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Adaptive SC-VAMP, a self-tuning extension of score-based vector approximate message passing. Each factor module re-estimates its own hyperparameters by a local tilted-EM M-step that reuses the tilted moments already computed by the SISO interface, leaving the VAMP message-passing structure unchanged. The authors prove, under standard large-system SE assumptions and an identifiability condition, that the true parameters together with the matched SE variance form a fixed point of the population recursion (Proposition 1), with the posterior estimate reaching the replica Bayes-optimal MMSE at that fixed point. The proof is split into prior-module and likelihood/LMMSE-module cases. Numerical experiments on linear and one-bit Bernoulli-Gaussian compressed sensing show that the adaptive method recovers near-oracle NMSE from mismatched initializations, while the frozen baseline does not.","tokens_in":1434,"tokens_out":1407,"duration_ms":46820,"significance":"If the self-tuning property is fully established, the paper contributes a modular and computationally cheap way to remove hyperparameter mismatch in VAMP-style inference, which is practically important because state-evolution predictions and reconstruction quality degrade sharply under mismatched priors or noise levels. The prior-module fixed-point argument is clean and general: it reduces the M-step to a KL divergence and is correct under the stated assumptions. The reuse of tilted moments from the SISO interface is elegant and keeps the per-module overhead at O(N). The numerical results are consistent and show a substantial improvement over the frozen baseline across both linear and one-bit settings. However, the theoretical claim is limited to a fixed-point statement, and the likelihood-module part rests on an unverified local identifiability condition; these gaps matter for the central 'self-tuning' claim.","major_comments":[{"comment":"Proposition 1 establishes that (theta0, vstar) is a fixed point of the population map (theta,v) mapping to (M(theta,v), F(v;theta)), but the paper's title and abstract claim 'self-tuning'. A fixed point is not convergence: the adaptive recursion could in principle start far from (theta0,vstar) and fail to approach it, or approach a different attractor. No local stability analysis, contraction argument, or convergence theorem is provided for the coupled (theta,v) recursion. The numerical experiments in Section III are encouraging but cover only two model families and a limited set of initializations; they do not replace an analysis of the population dynamics. This is load-bearing because the 'self-tuning' claim is exactly a statement about the behavior from mismatched initializations. The authors should either prove local stability of the fixed point (for example, by analyzing the Jacobia","section":"Section II-D, Proposition 1"},{"comment":"The proof for likelihood/LMMSE modules depends on a 'local compatibility condition' that is assumed rather than derived. Specifically, the M-step is analyzed under the Gaussian-cavity law p0_m(u_m,y_m|r_m)=gstar_m(u_m|r_m)phi_{theta0}(y_m|u_m), and the conclusion that equality in (23) implies theta_m=theta0_m requires that the cavity-weighted conditional KL identifies theta0_m. This condition is not verified for the probit likelihood used in the one-bit experiments. The paper itself notes (Section III-B) that above input SNR 10 dB the sign map becomes nearly scale-invariant and sigma_x^2 is not reliably identifiable; at that boundary the assumptions of Proposition 1 are not met. Thus the theorem does not cover part of the parameter regime in which the method is claimed to work. The authors should characterize when local compatibility holds (for example, for the probit channel below some","section":"Section II-D, likelihood-module proof and Eq. (22)-(23)"}],"minor_comments":[{"comment":"The notation 'V AMP' is inconsistently spaced; use 'VAMP' throughout.","section":"General"},{"comment":"The numerical experiments do not compare against existing EM-based VAMP methods such as EM-VAMP or EM-GVAMP. Such baselines would strengthen the claim that the proposed tilted-EM step is competitive, though their omission does not affect the paper's internal consistency.","section":"Section III"},{"comment":"The LMMSE noise update for sigma_w^2 is written as an implicit equation. It would be clearer to present the explicit fixed-point update, solving for the new sigma_w^2, to avoid ambiguity.","section":"Section II-C, Eq. (9)"},{"comment":"The one-bit SNR sweep is stopped at 10 dB because sigma_x^2 becomes non-identifiable. This is an honest limitation, but the sentence should appear earlier in the section so the reader knows the range is intentionally limited.","section":"Section III-B"}],"recommendation":"major_revision","confidential_remarks":"The fixed-point result is sound and the experiments are promising, but the gap between 'fixed point' and 'self-tuning' is substantial, and the likelihood proof relies on an identifiability condition that is known to fail in one of the two headline experiments. I think the paper can be made publishable by adding a stability analysis of the population recursion (at least for the linear-Gaussian case) or by substantially softening the self-tuning claims and clearly delimiting the theoretical scope. Given the current framing, major revision is appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper does something real — local tilted-EM M-steps that drop into SC-VAMP's existing SISO interface, so each factor updates its own hyperparameters without touching the Onsager correction. That is new relative to EM-GAMP/EM-VAMP/EM-GVAMP, which tune a fixed architecture or a specific module pair. The cost is O(N) per module because the tilted moments are already computed. The numerical story is consistent: adaptive SC-VAMP tracks the oracle curve in both linear and one-bit Bernoulli-Gaussian compressed sensing, and the paper is honest about the one-bit identifiability boundary around 10 dB input SNR.\n\nThe fixed-point theorem (Prop. 1) is the real contribution and it is largely clean. For prior modules the argument is a standard EM/KL calculation and it works. For likelihood modules the proof goes through only under a 'local compatibility condition' — the cavity-weighted conditional KL identifies the true parameter — and that condition is asserted, not derived. The paper itself flags this. It also only proves that (θ0, v*) is a fixed point of the population map, not that the damped recursion converges to it from mismatch. The numerics suggest convergence in these experiments, but there is no stability analysis. So the title's 'self-tuning' is supported by simulation, not by the theorem.\n\nThe soft spots are proportionate. Missing code and missing baselines against existing EM-AMP methods are annoying but fixable. The bigger issue is the gap between fixed point and convergence: an unstable fixed point would sink the practical claim even though Prop. 1 stays true. The one-bit experiment actually shows the identifiability limit, which is good science but also confirms the likelihood-module assumption is not generic.\n\nWho should read it: people working on AMP/VAMP parameter adaptation, or on compressed sensing with unknown sparsity/noise levels. It deserves a serious referee — the modular EM idea is worth having in the literature. For peer review I'd ask for convergence/stability analysis (even local), verification or a sharper statement of the compatibility condition, a baseline comparison, and code. With those it could be a solid journal paper; without them it's a well-written conference contribution.","headline":"Useful and honest extension of EM tuning to SC-VAMP's SISO interface; the fixed-point theorem is sound but the self-tuning claim outruns the proof.","tokens_in":8108,"tokens_out":1666,"would_cite":false,"duration_ms":18487,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Adaptive SC-VAMP attaches a tilted-EM self-tuning step to every module and proves that the true hyperparameters form a Bayes-optimal fixed point.","keywords":["adaptive SC-VAMP","vector approximate message passing","tilted expectation-maximization","hyperparameter self-tuning","compressed sensing","one-bit quantization","Bernoulli-Gaussian prior","state evolution"],"falsifier":"Synthesize a likelihood family with two parameter values that give the same cavity-averaged conditional distribution, run adaptive SC-VAMP from a mismatched start, and check whether the parameter estimate converges to the non-true value while the NMSE appears to improve; or extend the one-bit SNR sweep beyond 10 dB and observe whether σ̂_x² separates from its true value as the sign map becomes scale-invariant.","tokens_in":7211,"feed_emoji":"⚙️","tokens_out":6610,"duration_ms":64840,"temperature":0.7,"pith_summary":"This paper tackles a practical weak spot of vector approximate message passing (SC-VAMP): the algorithm's accuracy and its state-evolution predictions depend on knowing the signal prior's hyperparameters and the noise level, which in real problems are unknown. It attaches a local expectation–maximization step to every parameterized module, reusing the tilted moments the module already computes, so sparsity rate, active variance, and noise variance are re-estimated on the fly from the current posterior. The paper proves that under standard large-system assumptions and an identifiability condition, the true hyperparameters together with the matched state-evolution variance are a fixed point of the adaptive recursion, and the posterior estimate at that point reaches the replica (Bayes-optimal) mean-squared error. In linear and one-bit Bernoulli–Gaussian compressed sensing experiments, the adaptive method recovers near-oracle performance from strongly mismatched initializations, while leaving the message-passing interface unchanged.","feed_headline":"VAMP tunes its own hyperparameters: near-oracle compressed sensing","feed_subtitle":"Tilted EM updates recover true sparsity rates and noise levels from strongly mismatched starts.","key_machinery":"The mechanism is the local tilted-EM M-step: each module updates its factor parameters by maximizing E_{u∼p̃}[log φ(u,y;θ)], where p̃ is the tilted posterior formed by multiplying the factor with the Gaussian cavity message. Because the SISO (single-input single-output) interface already computes the needed tilted moments—the Tweedie posterior mean and the Fisher-information Onsager term—the update adds only O(N) work per module and leaves the score-based interface and Onsager correction untouched. For likelihood modules, the fixed-point proof uses the VAMP transformed-error Gaussian cavity law, which turns the M-step into a weighted conditional-KL minimization and thereby makes the true par","core_discovery":"The central claim is Proposition 1: under the paper's large-system assumptions (scalar-equivalent channel, population averaging, exact scores, unique replica fixed point) plus identifiability of each learned factor, the pair consisting of the true hyperparameters and the matched state-evolution variance is a fixed point of the population adaptive SC-VAMP recursion, and at that fixed point the posterior estimate attains the replica (Bayes) MMSE. For prior modules, the proof uses Gaussian symmetry of the cavity message to show that the population M-step objective differs from its value at the true parameter by a Kullback–Leibler divergence, which is nonnegative and zero only at the truth. For","pith_inferences":["The one-bit experiments stop at 10 dB input SNR because the sign map becomes nearly scale-invariant and σ_x² is not reliably identifiable; this suggests that a sharper characterization of the identifiability region—e.g., via the Fisher information of the tilted probit model—would tell users where self-tuning can be trusted and where it will silently fail.","The proof's 'local compatibility condition'—that the cavity-weighted conditional KL identifies θ₀—is an assumption that could be tested directly in simulation: choose a likelihood family where two parameter values give the same cavity-averaged likelihood, and check whether the adaptive fixed point drifts from the true value.","Since the tilted-EM update only needs the tilted moments, a natural extension is to learned-score priors where the explicit factor is unavailable; in that case the M-step would need to be replaced by a score-matching or moment-matching surrogate, and the fixed-point argument would have to be reworked.","The damping factor β is treated as a user choice, but the one-bit experiments show β=0.3 is needed for stability; a principled schedule or adaptive damping rule derived from the Jacobian of the population map (θ,v) ↦ (M(θ,v), F(v;θ)) is a testable extension."],"forward_implications":["Self-tuning hyperparameters inside SC-VAMP removes the need for a separate cross-validation or parameter-search loop in high-dimensional inverse problems, since each module updates its own parameters from the tilted moments it already has.","At the matched population fixed point, the recursion coincides with matched SC-VAMP/VAMP, so the state-evolution predictions remain valid and the posterior estimate is Bayes-optimal under the replica assumption; adaptive runs inherit the performance guarantees of the non-adaptive algorithm.","Both prior modules (e.g., Bernoulli–Gaussian sparsity rate and active variance) and likelihood/LMMSE modules (noise variance) are covered, making the method applicable to linear and nonlinear (one-bit quantized) observation models.","Numerical experiments show the adaptive method tracking the oracle NMSE across SNR ranges where the frozen baseline is 1.5–2.5× worse (linear) and up to 3.1× worse at high SNR (one-bit), so the practical gain grows exactly where prior mismatch hurts most.","The M-step reuses already-computed tilted moments, so per-module overhead is O(N), and the only added user choice is the damping factor β."],"fun_headline_variants":["VAMP learns its own priors and noise via tilted EM","Self-tuning VAMP: tilted EM finds oracle parameters","Adaptive SC-VAMP: self-tuned hyperparameters from tilted EM","VAMP auto-tunes via tilted EM: near-oracle CS from bad starts","Tilted EM makes VAMP self-tuning: near-oracle compressed sensing"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The proof rests on the local compatibility condition: each module's true hyperparameter must be the unique maximizer of the likelihood after averaging over the Gaussian cavity; when that fails, as with the signal scale in high-SNR one-bit observations, the self-tuning step cannot recover the truth.","fun_headline_variants_meta":{"raw":{"variants":["VAMP learns its own priors and noise via tilted EM","Self-tuning VAMP: tilted EM finds oracle parameters","Adaptive SC-VAMP: self-tuned hyperparameters from tilted EM","VAMP auto-tunes via tilted EM: near-oracle CS from bad starts","Tilted EM makes VAMP self-tuning: near-oracle compressed sensing"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000927,"raw_usage":{"total_tokens":3775,"prompt_tokens":675,"completion_tokens":3100,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":419,"completion_tokens_details":{"reasoning_tokens":3007}},"tokens_in":419,"tokens_out":3100,"duration_ms":19966,"temperature":1.0,"reasoning_tokens":3007,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T01:38:24.984871+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Synthesize a likelihood family with two parameter values that give the same cavity-averaged conditional distribution, run adaptive SC-VAMP from a mismatched start, and check whether the parameter estimate converges to the non-true value while the NMSE appears to improve; or extend the one-bit SNR sweep beyond 10 dB and observe whether σ̂_x² separates from its true value as the sign map becomes scale-invariant.","supporting_citations":[],"review_version":1}