{"id":"eb0ad345-1cc6-4862-a663-8b340e497810","arxiv_id":"2508.02117","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":4.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"Score-based generative models can estimate ISAC performance metrics such as MI, MMSE, and BCRB from data, matching analytical ground truth in proof-of-concept tests.","lead":"This paper explains how to estimate standard wireless performance limits, such as mutual information and Cramér-Rao bounds, by learning probability score functions from data instead of solving hard equations. It applies the idea to integrated sensing and communications (ISAC) and runs small proof-of-concept experiments that match analytical answers.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Score-to-metric error propagation is untested: small pointwise score errors can materially bias plugin MI/MMSE/BCRB estimates in exactly the multimodal regimes the paper claims to reach.","rationale":"The reader's weakest_assumption is close to my concern: both focus on learned-score accuracy in multimodal regimes. I agree with that and add specificity: the issue is not raw score accuracy but error transfer through nonlinear and inverse forms of the metrics, especially the Fisher information inversion used in BCRB. I do not claim the paper is wrong; I claim the evidence provided is not sufficient to rule out large error amplification. Because the full text is corrupted with mojibake, I could not inspect the derivations or the actual experimental conditions, so I cannot move the paper off the reader's UNVERDICTED status. The proof-of-concept validation, if readable, would count as supporting evidence, but the current submission does not make it possible to verify that this validation covers the non-Gaussian multimodal regime where the central claim is most consequential.","tokens_in":33990,"tokens_out":5719,"duration_ms":81802,"concrete_test":"On a two-component Gaussian mixture with known ground truth, including numerically integrated BCRB and MI, train score networks under the paper's reported protocol and sweep component separation (0.5, 2, 4, 6) and dimension (1, 2, 4, 8). Plot normalized estimation error of MI, MMSE, and BCRB against the trained network's L2 score error. If metric error grows steeply when score error is moderate, or if BCRB is biased upward even at moderate score error, the claimed extension to multimodal non-Gaussian regimes is not supported by the current evidence. If metric error tracks score error smoothly and remains small across the sweep, the concern is discharged.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that a learned score can be plugged into formulas for MI, MMSE, and BCRB and give accurate performance numbers in non-Gaussian, multimodal ISAC settings. The load-bearing premise is quantified error transfer: if ||s_theta - nabla log p|| is small in the score-matching loss, the metrics computed from s_theta are close to the true metrics. The paper's proof-of-concept validation cannot establish this. Score-matching losses are weighted L2 averages over the data distribution; they permit large pointwise errors in low-density regions, and those regions are precisely where posterior Fisher information and entropy-like MI integrands are sensitive. For BCRB, the estimator typically requires the outer product of the posterior score or its Jacobian; squared score errors enter additively, likely inflating the Fisher matrix, while derivative errors of a neural network are typically larger than score errors. The abstract offers no error-propagation bound, no experiment that varies score error and observes metric error, and no test in a demonstrably multimodal, non-Gaussian high-dimensional example. Agreement with ground truth on proof-of-concept cases with analytic posteriors is consistent with a method that works there but does not resolve the target regime.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a framework, called \"scoring ISAC,\" in which score-based generative models are used to estimate classical ISAC performance metrics—mutual information (MI), minimum mean squared error (MMSE), and the Bayesian Cramér–Rao bound (BCRB)—from data, with the goal of handling nonlinear, multimodal, non-Gaussian settings where closed-form expressions are unavailable. The abstract claims a tutorial-style synthesis of recent advances, explicit connections between performance metrics and score functions, practical training techniques, and proof-of-concept experiments on target detection and localization that validate the score-based estimators against ground-truth analytical expressions. The delivered full text, however, is largely corrupted and unreadable, so the derivations, experimental details, and numerical results cannot be independently assessed.","tokens_in":34200,"tokens_out":2689,"duration_ms":33748,"significance":"If the claims are correct, the framework offers a genuinely useful data-driven route to performance evaluation in realistic ISAC scenarios, and the tutorial-style organization could benefit a broad audience. The paper also has the merit of connecting established information-theoretic identities to modern generative modeling in a new application domain. That said, the central claim—that plugin estimates from a learned score are accurate in exactly the multimodal, non-Gaussian regimes of interest—is not verifiable from the submitted manuscript, because the full text is unreadable. No machine-checkable proofs, reproducible code, or parameter-free derivations are visible in the delivered file. The significance can be evaluated only after a readable resubmission.","major_comments":[{"comment":"The submitted text is almost entirely corrupted: most sentences are unreadable mojibake, the running header cites arXiv:2508.02120 rather than 2508.02117, and large blocks of the text, including the derivations and experimental sections, are indecipherable. This prevents verification of every load-bearing element of the paper—the score-to-metric identities, the training procedures, and the proof-of-concept validation. The authors must resubmit a readable manuscript (PDF or LaTeX source) before the scientific content can be reviewed.","section":"Full text (as delivered); title page/header"},{"comment":"The abstract states that proof-of-concept experiments \"validate the accuracy of score-based performance estimators against ground-truth analytical expressions,\" but the experiments section is unreadable, so this validation cannot be checked. Moreover, the abstract does not clarify whether the training data are generated from the same statistical model that supplies the ground-truth analytical expressions; if that is the case, the agreement would largely be a self-consistency check and would not demonstrate accuracy in the target non-Gaussian, multimodal regimes. The resubmission should state the data-generation protocol explicitly and, if the same model is used, discuss why the test is not circular.","section":"Abstract, last sentence"}],"minor_comments":[{"comment":"The LaTeX macro \"Cram\\'{e}r--Rao bound\" appears with a stray backslash; this should be cleaned up.","section":"Abstract"},{"comment":"The arXiv identifier in the running header (2508.02120) does not match the submitted paper's identifier (2508.02117); this should be corrected.","section":"Header/title page"},{"comment":"The reference list at the end of the document is garbled and appears to contain repeated or partial entries; it needs to be regenerated from the original source files.","section":"Reference list"},{"comment":"The few readable fragments suggest the document contains figures and tables, but their captions and contents are not legible; the resubmission should ensure all figures and tables render properly.","section":"Figures/tables (where readable)"}],"recommendation":"major_revision","confidential_remarks":"The corruption of the full text is severe enough that no substantive technical evaluation is possible at this stage. I recommend sending the manuscript back for a readable resubmission before any further review. I also note that the mismatch between the arXiv ID in the header and the submitted ID may indicate a source/pipeline error; if so, the authors should be alerted so they can supply the correct file."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a first-application of score-based generative modeling to ISAC performance metrics, framed as a tutorial plus proof-of-concept. The abstract is clear and the motivation is sound: closed-form MI/MMSE/BCRB fail in non-Gaussian multimodal settings, and score functions offer a data-driven route. If the full paper delivers what the abstract promises, it will be useful to the ISAC community.\n\nWhat is actually new: applying existing score-to-metric connections to ISAC, with target detection and localization examples. The authors don't overclaim; they explicitly say it has not been discussed for ISAC. That is honest. The proof-of-concept against analytical ground truth is a reasonable sanity check, though it mainly confirms self-consistency when training data come from the same model that supplies the ground truth.\n\nSoft spots. First, the full text I received is mojibake; I can only judge the abstract. Any verdict on derivations, experiment details, or error bars is impossible, and that alone forces a cautious reading. Second, the central premise is unquantified: small score error in a weighted L2 sense does not obviously imply small error in MI, MMSE, or BCRB, because low-density regions matter most and error there is not controlled. The abstract offers no error-propagation bound and no experiment that varies score error and observes metric error. That is a real gap, though it may be addressed in the corrupted sections. Third, the validation cases use analytic posteriors; they do not by themselves demonstrate the multimodal non-Gaussian regime the paper's motivation rests on.\n\nCitation pattern: I cannot check it from the text. No red flags from the abstract.\n\nWho this is for: ISAC researchers who want a practical recipe for estimating otherwise intractable performance metrics, and people working on score-based estimation who want a new application domain. It deserves a serious referee, but only after the authors supply a readable PDF. My recommendation: ask for a clean version and send it out; the core idea is plausible and useful enough to merit careful checking.","headline":"A tutorial-style transfer of score-based metric estimation to ISAC with proof-of-concept checks; worth refereeing, but the text as delivered is unreadable and the error-propagation question is unaddressed in the abstract.","tokens_in":34700,"tokens_out":1818,"would_cite":false,"duration_ms":24255,"reading_group":"maybe","serious_thinker":"unclear","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Score-based generative models can estimate ISAC performance metrics—mutual information, MMSE, and Bayesian Cramér–Rao bound—directly from data, with proof-of-concept experiments matching analytical ground truth.","keywords":["integrated sensing and communications","score-based generative models","mutual information","minimum mean squared error","Bayesian Cramér-Rao bound","data-driven performance evaluation","non-Gaussian signal processing","target localization"],"falsifier":"Run the estimator on a synthetic ISAC observation model with known ground truth—for example, a Gaussian mixture or a nonlinear localization model whose mutual information and Bayesian Cramér–Rao bound can be computed numerically—and check whether the score-based estimates stay accurate as the dimension grows and the distribution becomes more multimodal; growing discrepancy would falsify the central claim.","tokens_in":33809,"feed_emoji":"📡","tokens_out":3695,"duration_ms":45223,"temperature":0.7,"pith_summary":"This paper is a tutorial-style proposal for scoring ISAC: use score-based generative models to evaluate integrated sensing and communication systems directly from data. The central claim is that three classical performance metrics—mutual information, minimum mean squared error, and the Bayesian Cramér–Rao bound—can be expressed in terms of score functions, so once a model learns the score of the relevant observation or posterior density, all three metrics become available even when the underlying distributions are nonlinear, multimodal, and non-Gaussian. Proof-of-concept experiments on target detection and localization compare the learned estimates with ground-truth analytical expressions and report agreement. If the claim holds, ISAC performance evaluation no longer requires closed-form statistical models and can follow the actual data distribution instead.","feed_headline":"Score models read ISAC performance from raw data","feed_subtitle":"Learned score functions yield mutual information, MMSE, and Bayesian bounds, matching analytic baselines in detection and localization.","key_machinery":"The load-bearing object is the learned score function $\\mathbf{s}_{\\boldsymbol{\\theta}}(\\mathbf{x}) \\approx \\nabla_{\\mathbf{x}}\\log p(\\mathbf{x})$, the gradient of the log-density of the data. The paper establishes score-based expressions for mutual information, MMSE, and the Bayesian Cramér–Rao bound, and uses denoising score matching to train the network so that the same learned score can be plugged into all three expressions. That shared score map is what carries the argument: once it is accurate, the classical metrics follow without any new analytic derivation.","core_discovery":"On its own terms, the paper claims that score functions carry enough statistical information to serve as a universal performance evaluator for ISAC systems. It argues that the score $\\nabla_{\\mathbf{x}}\\log p(\\mathbf{x})$ of the joint or conditional density encodes the same information that enters mutual information, MMSE, and the Bayesian Cramér–Rao bound, and that standard score-matching training therefore turns a generative model into a plugin estimator for all three metrics. The validation experiments show that this procedure reproduces closed-form baselines in detection and localization, which the paper presents as evidence that the mechanism extends to realistic regimes where analytical derivations are unavailable.","pith_inferences":["A natural next stress test is dimensionality: in massive-antenna or multi-target settings, score estimation error will compound through the metric identities, and the paper does not characterize this degradation.","The framework implicitly separates score-learning error from plugin-evaluation error; isolating the two would clarify when the method can be trusted.","Because the metric estimates are differentiable functions of the learned score, they could be differentiated to guide waveform or beamforming optimization, an extension the paper gestures toward but does not develop."],"forward_implications":["ISAC system evaluation becomes a data-driven procedure: collect samples from the noisy, impaired channel and estimate mutual information, MMSE, or the Bayesian Cramér–Rao bound from the learned score.","One trained score model supplies all three metrics at once, instead of requiring a separate analytical treatment for each.","Detection and localization limits can be quantified in non-Gaussian, multimodal scenes where closed-form Bayesian bounds do not exist.","The same estimated metrics can be used as objectives or constraints for ISAC algorithm design and system optimization."],"supporting_citations":[],"fun_headline_variants":["Score models match ISAC analytic baselines","Score functions estimate ISAC metrics from data","Scoring ISAC via score-based generative models","Score-based estimators benchmark ISAC performance"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole method rests on the learned score function being a close approximation to the true log-density gradient in exactly the non-Gaussian, multimodal settings where closed-form analysis fails; the paper validates this on proof-of-concept cases but states no general approximation guarantee.","fun_headline_variants_meta":{"raw":{"variants":["Score models match ISAC analytic baselines","Score functions estimate ISAC metrics from data","Scoring ISAC via score-based generative models","Score-based estimators benchmark ISAC performance"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001084,"raw_usage":{"total_tokens":4537,"prompt_tokens":955,"completion_tokens":3582,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":571,"completion_tokens_details":{"reasoning_tokens":3527}},"tokens_in":571,"tokens_out":3582,"duration_ms":27548,"temperature":1.0,"reasoning_tokens":3527,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T05:07:40.331489+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the estimator on a synthetic ISAC observation model with known ground truth—for example, a Gaussian mixture or a nonlinear localization model whose mutual information and Bayesian Cramér–Rao bound can be computed numerically—and check whether the score-based estimates stay accurate as the dimension grows and the distribution becomes more multimodal; growing discrepancy would falsify the central claim.","supporting_citations":[],"review_version":1}