{"id":"7ece8db4-a686-46d7-bb57-43c981d3f7c9","arxiv_id":"2412.09839","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"This review proposes that model-driven deep learning, CSI foundation models, and LLMs can address the compute, modeling, and measurement challenges of terahertz ultra-massive MIMO.","lead":"This paper lays out three research roadmaps for applying AI to terahertz ultra-massive MIMO: model-driven deep learning, a proposed channel-focused 'CSI foundation model', and large language models. It is a review and roadmap, not a new experimental result; a smart generalist might read it for a map of where 6G physical-layer AI research is heading.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The raw-signal training path in §4.2.1 is the weakest load-bearing step: SURE/GSURE give the score of noisy measurements at one noise level, not the clean-channel score across the diffusion schedule needed for a generative prior.","rationale":"The manuscript is a roadmap/review, so its central claim is a proposal rather than a demonstrated result. That lowers the bar: a conditional acceptance is appropriate if the proposal is coherent and actionable. The reader's weakest-assumption pick, low effective dimensionality of THz channels, is a genuine empirical risk; I agree that the paper's compactness and sample-efficiency arguments lean on it. My pass, however, found a more specific and more decisive gap in the training-from-raw-signals leg of the proposal. Section 4.2.1 explicitly says the score of clean channels can be estimated from y via SURE/GSURE. Those techniques estimate the score of the observed noisy density or train a denoiser at the measurement noise level. They do not provide scores at noise levels below σ², which are required to generate clean channels by reverse diffusion. Without a separate mechanism (deconvolution, multiple noisy views per channel, a parametric channel family, or access to some clean data), the 'hard-to-measure' path cannot produce the open-ended generative prior that the abstract promises. The case study is honest: it uses the denoiser for channel estimation, which works, but it does not demonstrate prior sampling, compression, or joint sampling from a model trained this way. This concern is partly orthogonal to the low-dimensionality issue: even a perfectly low-dimensional channel distribution is not recoverable from a single noisy observation per sample by SURE alone. I therefore recommend keeping the reader's conditional verdict. The paper should be accepted only after the authors either correct the raw-signal training claim, add the missing identifiability/regularity assumptions, or reframe the roadmap so that the generative foundation model is trained on clean channel data (when available) with the SURE-based denoiser presented as a downstream module rather than as the foundation model itself. I credit the paper for being explicitly forward-looking and for including an initial case study, but that case study does not settle the raw-signal training question.","tokens_in":32493,"tokens_out":9041,"duration_ms":102814,"concrete_test":"Use a known synthetic prior, e.g., a Gaussian mixture on h∈R^64, and generate y=h+n with n~N(0,σ²I). Train a score network on y only using the SURE-loss procedure described in §4.2.1, then run reverse diffusion with the learned scores. Compare the resulting samples to the true p(h) via a two-sample test or density estimate (e.g., MMD on summary statistics). If the samples reproduce p(y) but fail to reproduce p(h), the raw-signal training claim is false; if they reproduce p(h), the claim survives. Supplement with the analytical check of whether deconvolution, F[p(y)]/exp(-σ²||ω||²/2), is numerically stable at the noise levels used.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.2.1 claims that from raw received signals y=h+n, 'the (prior) score function ∇_h log p(h) can be estimated using empirical Bayesian methods based on the SURE loss' and extends this to GSURE for y=Mh+n. This is the pathway that carries the paper's hard-to-measure motivation: clean channel data are assumed scarce, so the foundation model must be trainable from corrupted measurements. The technical problem is that SURE/GSURE and Stein-based explicit score matching estimate the score of the noisy measurement density, ∇_y log p(y), not the clean-channel score ∇_h log p(h). At a fixed noise variance σ², Tweedie's formula turns that score into the MMSE denoiser E[h|y], which is useful for estimation but not sufficient for a generative prior. Score-based generation requires scores across a noise schedule down to t→0; adding synthetic noise to y gives scores only at total noise levels ≥ σ². The low-noise levels that produce clean channel samples for data augmentation and for sequential/joint sampling in §4.2.1 are not identifiable from one noisy observation per channel without deconvolution or additional assumptions. The case study in §4.2.5 indeed trains only a denoiser for y=h+n, which is consistent with this limitation but does not yet demonstrate the promised compact generative foundation model. This is an internal gap in the roadmap, not merely a matter of empirical uncertainty: as written, the raw-signal path does not deliver the object it defines.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper is a research-roadmap review for applying AI and deep learning to terahertz ultra-massive MIMO (UM-MIMO) systems. It identifies three challenges, termed \"hard-to-compute,\" \"hard-to-model,\" and \"hard-to-measure,\" and proposes three roadmaps: (1) model-driven deep learning based on fixed point networks and neural calibration, with steps for algorithmic frameworks, basis algorithms, loss design, and neural architecture; (2) a \"CSI foundation model\" that learns the score function of wireless channels and uses it as a shared generative prior for estimation, compression, detection, and other transceiver tasks; and (3) applications of large language models to estimation, optimization, search, network management, and protocol understanding. The paper includes two model-driven case studies (beam-focusing and data detection) and one CSI foundation model case study (channel estimation), all reproduced from the authors' prior peer-reviewed work. The central original proposal is that a single compact score-based generative model, trained on either clean channels or raw received signals, can serve as an open-ended prior for many physical-layer tasks.","tokens_in":1754,"tokens_out":2121,"duration_ms":52521,"significance":"If the CSI foundation model concept works as envisioned, it would be a substantial step toward unifying transceiver design in 6G systems: a single learned channel prior could replace task-specific networks, reducing training and deployment cost. The paper also provides a useful synthesis of model-driven DL techniques, including fixed point networks, neural calibration, SURE/GSURE losses, graph neural networks, hypernetworks, transform-domain learning, and NeRF2/Gaussian splatting, and connects them to terahertz-specific channel properties such as the hybrid-field effect, beam squint, and array-of-subarrays architecture. The explicitly identified open problems in Section 5 are honest about the scarcity of real terahertz measurements and the unresolved scalability issues. However, the paper makes no new derivations, and its case-study evidence is entirely reproduced from prior simulations; the load-bearing novelty is the roadmap for training a generative prior from corrupted measurements, and that roadmap currently has an internal technical gap.","major_comments":[{"comment":"The raw-signal training pathway asserts: \"the (prior) score function ∇_h log p(h) can be estimated using empirical Bayesian methods based on the SURE loss\" for y = h + n and via GSURE for y = Mh + n. This conflates the score of the noisy measurement density with the clean-channel score. SURE/GSURE and Stein-type unbiased risk estimators provide, at a fixed noise variance, the MMSE denoiser E[h|y] via Tweedie's formula; they do not provide ∇_h log p(h) across the noise schedule t → 0 that a score-based generative prior requires. Adding synthetic noise to y only yields scores at total noise levels greater than or equal to the observation noise level, so the low-noise portion of the diffusion schedule needed for generation and joint sampling is not identifiable from one noisy observation per channel without additional deconvolution or modeling assumptions. This is a load-bearing gap because the hard-to-measure motivation of the CSI foundation model depends on training from corrupted measurements, and the case study in Section 4.2.5 trains only a denoiser, not a generative score model. The authors should either weaken the claim to \"MMSE denoiser at the operating noise level\" or provide a concrete method for obtaining the full score schedule from noisy data.","section":"Section 4.2.1"},{"comment":"The case study presented as a CSI foundation model demonstration trains a denoising auto-encoder on y = h + n and then uses Tweedie's formula to obtain an MMSE denoiser for the OAMP framework. This validates a denoiser-based plug-and-play prior at a fixed noise level, but it does not demonstrate the generative capabilities central to the roadmap: prior sampling for data augmentation, posterior sampling for inverse problems, sequential sampling for compression, or joint sampling for cross-module design are not implemented or evaluated. The text states \"we train the score network s_θ(h) by using denoising auto-encoders based on y [94]\" and then immediately derives a denoiser; there is no evidence that the learned network can be sampled as a score-based generative model. The authors should explicitly position this case study as a denoiser-only instantiation and identify which parts of the roadmap remain unvalidated.","section":"Section 4.2.5"},{"comment":"The compactness and sample-efficiency of the CSI foundation model rest on the premise that \"the effective dimensionality of high-frequency wireless channels is small because of the limited number of paths\" (Section 4.2.1). The paper cites a measurement at 300 GHz with five multipaths (Section 2.2.2), but it simultaneously emphasizes hybrid-field effects, spatial non-stationarity, and wideband beam squint, which complicate sparse representations and may increase the intrinsic complexity of the channel distribution. The roadmap would be strengthened by formulating this as a falsifiable hypothesis, for example by specifying how the sample complexity of score matching scales with the number of paths and array size, rather than presenting the low-dimensionality claim as an established fact. As written, the reader cannot tell whether the \"compact\" foundation model is a design principle or an empirical conjecture.","section":"Section 2.2.2 and Section 4.2.1"}],"minor_comments":[{"comment":"The subsection numbering is duplicated: both \"Conditioning\" and \"Site-specific adaptation\" are numbered 4.2.2. The second should be 4.2.3.","section":"Section 4.2"},{"comment":"The abstract states \"Four essential steps are outlined: general frameworks, conditioning, site-specific adaptation, joint design of CSI foundation models, and model-driven DL,\" but this lists five items. The list should be reduced to four, as in Section 4.2, or the sentence should be rephrased.","section":"Abstract"},{"comment":"The caption defines the under-sampling ratio as \"the number of columns in measurement matrix M divided by its number of rows,\" which describes an over-sampling ratio for a fat measurement matrix. If the intended quantity is the usual compressive-sensing ratio M/N (rows/columns), the definition should be corrected for consistency with the literature.","section":"Fig. 8 caption"},{"comment":"Reference [154] is missing from the reference list, and the in-text citation sequence jumps from [153] to [155] in Section 4.2. The numbering should be corrected.","section":"References"},{"comment":"In the sentence \"we discuss the manifestation of these challenges in terahertz UM-MIMO systems, this explains the motivation for integrating DL with terahertz UM-MIMO systems,\" a comma splice is used; it should be rephrased as two sentences or joined with a semicolon.","section":"Section 3"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. This is a roadmap/review paper, not a research result: no new algorithms, no new measurements. What it does well is structure a young field—model-driven deep learning, score-based CSI foundation models, and LLM applications for terahertz UM-MIMO—into concrete, ordered steps that a newcomer can actually follow. The 'CSI foundation model' framing is genuinely useful, and the authors openly acknowledge a concurrent paper on a similar concept (Ref. [155]) and position their contribution as complementary. That is honest.\n\nThe soft spot is the load-bearing piece of Roadmap 2. In §4.2.1 the paper says that when only raw received signals y=h+n or y=Mh+n are available, the clean-channel score ∇_h log p(h) can be estimated with SURE/GSURE. That is not what SURE/GSURE give you. They give an unbiased estimate of the risk of a denoiser, and by Tweedie's formula the learned score is that of the noisy density ∇_y log p(y), not the clean density. Score-based generation needs scores across a noise schedule all the way down to t→0; with a single noisy observation per channel you only have access to noise levels ≥ σ². The case study in §4.2.5 actually trains a denoiser, uses Tweedie's formula, and plugs it into OAMP as an MMSE denoiser—that is a coherent, useful result, but it is not a generative foundation model from raw measurements. The paper does not flag this gap; it switches from the raw-signal claim to the denoiser case study without noting the discrepancy.\n\nEverything else is more solid. The model-driven DL roadmap is well grounded in prior work, the literature coverage is broad and current, and the reproduced case studies are from peer-reviewed papers by the same group. Minor annoyances: duplicate subsection number 4.2.2, a corrupted Rayleigh distance formula in Section 2.2.1, and the usual roadmap-paper problem that the simulation figures are replotted without error bars. Those are fixable in revision.\n\nWho should read this: anyone entering AI-for-terahertz research, and people in the foundation-model-for-wireless community who want a structured map of the design space. It deserves a serious referee: the roadmap is useful, but the raw-signal training claim needs to be either corrected or carefully scoped. If the authors revise §4.2.1 to state that SURE/GSURE give a denoiser for the noisy measurements and that a full generative model requires either clean data or a noise-schedule-aware training procedure, the paper would be on firm ground. Send it to review.","headline":"A useful roadmap review whose central raw-signal training claim for a CSI foundation model has a real technical gap; the rest is solid and it deserves peer review.","tokens_in":33320,"tokens_out":2726,"would_cite":false,"duration_ms":28686,"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 single compact neural network that learns the score function of terahertz UM-MIMO wireless channels could serve as a shared foundation for channel estimation, data detection, CSI compression, and data augmentation, replacing many…","keywords":["terahertz communications","ultra-massive MIMO","model-driven deep learning","CSI foundation models","score-based generative models","channel estimation","6G","large language models"],"falsifier":"A measurement campaign at a terahertz UM-MIMO testbed that counts resolvable multipaths as array aperture and bandwidth scale: if the number of significant paths grows with system scale or environment complexity so that no compact score model matches the empirical channel distribution, the foundation-model premise fails. A simpler experiment: train the proposed score network on raw pilots from a measured indoor terahertz channel and compare channel-estimation NMSE against least-squares and sample-MMSE estimators; if the score-based prior does not beat these baselines across SNR, the central claim is not supported.","tokens_in":32262,"feed_emoji":"📶","tokens_out":6149,"duration_ms":59866,"temperature":0.7,"pith_summary":"This paper argues that the many separate neural networks used in terahertz ultra-massive MIMO transceivers are redundant because they all depend on the same underlying object, the wireless channel. Its central proposal is a 'CSI foundation model': one compact generative model trained to estimate the score function of the channel distribution, which then acts as a reusable prior for channel estimation, data detection, CSI compression, data augmentation, and other physical-layer tasks. The paper lays out three research roadmaps—model-driven deep learning, CSI foundation models, and large-language-model applications—and treats the foundation-model roadmap as the unifying vision. If the proposal holds, deployment cost and training redundancy would drop while transceiver modules would share a single learned representation of the propagation environment.","feed_headline":"Score-based channel prior could unify terahertz 6G transceivers","feed_subtitle":"A compact model that learns the channel score function could replace dozens of task-specific neural networks in 6G.","key_machinery":"The load-bearing object is the score network, $\\mathbf{s}_\\theta(\\mathbf{h}) \\approx \\nabla_{\\mathbf{h}} \\log p(\\mathbf{h})$, a neural estimator of the channel score function. The score function is connected to the MMSE denoiser through Tweedie's formula, which lets the same network serve as the non-linear estimator inside OAMP and other approximate-message-passing algorithms; this connection is what makes the prior open-ended. Training uses denoising score matching when clean channels exist and SURE or GSURE losses when only raw received signals $\\mathbf{y} = \\mathbf{M}\\mathbf{h} + \\mathbf{n}$ are available. The paper also relies on fixed-point networks, which turn iterative algorithms into contractive learned operators with convergence guarantees, as the interface for plugging the score prior into model-driven deep learning.","core_discovery":"The central discovery, proposed rather than experimentally established, is that the score function of terahertz UM-MIMO channels can be learned once and reused everywhere. The score function is the gradient of the log-density of the channel distribution, a vector field pointing toward more probable channel states. Trained by denoising score matching on clean channel data or, via SURE and GSURE losses, directly on raw noisy and incomplete received pilots, the score network becomes an open-ended Bayesian prior. With it, a designer can perform prior sampling to synthesize channels, posterior sampling to estimate channels from measurements, sequential sampling to compress and reconstruct CSI, and joint sampling for tasks such as combined channel estimation and data detection. The paper demonstrates the concept in a near-field 1024-antenna UM-MIMO channel estimation case study, where the score-based prior plugged into an OAMP fixed-point network approaches the oracle MMSE bound.","pith_inferences":["The paper leaves implicit that the same score prior could extend beyond transceiver modules to joint sensing and communication, localization, and environment mapping, since those tasks also depend on the channel distribution; the paper's non-exhaustive sampling list invites this extension.","A testable near-term extension is to pre-train the score network on synthetic channels generated by a site-specific ray tracer, then fine-tune it on raw received pilots using low-rank adaptation; this would test whether measurement burden can be reduced further than the paper's case study.","If the compactness claim holds, CSI foundation models could be shared across base stations through federated low-rank adaptation, preserving privacy while adapting to each site; the paper mentions federated LoRA but does not develop this for terahertz UM-MIMO."],"forward_implications":["A single pretrained score network replaces task-specific channel estimation, detection, and compression networks, cutting training and deployment cost.","Transceiver modules can be designed without clean channel labels: SURE and GSURE losses allow training from raw received pilots, which matters when ground-truth channels are hard to measure.","Model-driven deep learning and foundation models compose: the score prior plugs into OAMP and AMP as an MMSE denoiser, so convergence-guaranteed algorithms can use a learned prior.","Data augmentation from the prior reduces the need for large measurement campaigns, directly addressing the hard-to-measure problem.","Conditional score networks using labels such as weather, line-of-sight status, or position let one model serve many sites and scenarios without retraining."],"supporting_citations":[{"why":"Supplies the score-based generative modeling framework that defines how a learned score function can generate samples and act as a prior.","marker":"[152]"},{"why":"The paper's own case study training a score network from raw received pilots and applying it to OAMP for near-field UM-MIMO channel estimation.","marker":"[94]"},{"why":"Provides the fixed-point network framework and the terahertz UM-MIMO channel model that the case studies build on.","marker":"[38]"},{"why":"Connects denoisers to AMP-family algorithms, establishing the plug-and-play route through which the score prior enters iterative transceiver algorithms.","marker":"[124]"},{"why":"Generalized SURE enables score training from raw incomplete measurements, which is essential under the array-of-subarrays architecture with few RF chains.","marker":"[130]"},{"why":"Denoising score matching is the core training method for the score network when clean channel data are available.","marker":"[156]"},{"why":"Demonstrates score-based data augmentation for wireless channels, supporting the prior-sampling application of the CSI foundation model.","marker":"[91]"},{"why":"Low-rank adaptation is the parameter-efficient fine-tuning method proposed for site-specific adaptation of the foundation model.","marker":"[176]"}],"fun_headline_variants":["Score function as universal prior for 6G MIMO","One score network to replace task-specific 6G nets","Proposed score prior could unify terahertz transceivers","Learn channel score once, apply to all 6G tasks","A single score model for many terahertz MIMO modules"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Terahertz UM-MIMO channels can be captured by a compact generative model because they have few significant multipaths, making their effective dimension much smaller than the antenna count; if real channels are not that low-dimensional, the score prior will be too weak to support the promised transceiver modules.","fun_headline_variants_meta":{"raw":{"variants":["Score function as universal prior for 6G MIMO","One score network to replace task-specific 6G nets","Proposed score prior could unify terahertz transceivers","Learn channel score once, apply to all 6G tasks","A single score model for many terahertz MIMO modules"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000357,"raw_usage":{"total_tokens":1981,"prompt_tokens":1034,"completion_tokens":947,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":650,"completion_tokens_details":{"reasoning_tokens":862}},"tokens_in":650,"tokens_out":947,"duration_ms":10319,"temperature":1.0,"reasoning_tokens":862,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T16:39:55.630391+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A measurement campaign at a terahertz UM-MIMO testbed that counts resolvable multipaths as array aperture and bandwidth scale: if the number of significant paths grows with system scale or environment complexity so that no compact score model matches the empirical channel distribution, the foundation-model premise fails. A simpler experiment: train the proposed score network on raw pilots from a measured indoor terahertz channel and compare channel-estimation NMSE against least-squares and sample-MMSE estimators; if the score-based prior does not beat these baselines across SNR, the central claim is not supported.","supporting_citations":[{"cited_title":"Generalized SURE for exponential families: applications to regularization","cited_arxiv_id":null,"evidence_quote":"Generalized SURE enables score training from raw incomplete measurements, which is essential under the array-of-subarrays architecture with few RF chains."}],"review_version":1}