{"id":"b2acd179-a12a-4de5-8fe5-eeaa1d3aa241","arxiv_id":"2412.18162","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"An unsupervised teacher-student deep learning approach jointly designs communication and sensing beamformers for cell-free ISAC, achieving near-CVX performance with a reported three-order-of-magnitude runtime reduction.","lead":"This paper trains neural networks to design beamforming vectors for cell-free systems that communicate with users and sense a target at the same time, using a teacher-student scheme to balance the two goals. The authors report performance close to a standard optimization solver while running three orders of magnitude faster.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper never quantifies the gap caused by letting each AP's DNN see only local CSI, even though the min-SINR objective couples all APs; this unvalidated approximation underpins the distributed/fronthaul claim.","rationale":"The reader's weakest-assumption analysis identifies the same structural issue: the distributed DNNs use only local CSI for an objective that is globally coupled. My reading of the system model confirms this. Eq. (2) makes the coupling explicit: SINR at user n involves h_n^H w_n, where h_n and w_n are stacked across all APs, so the contribution of AP l to the useful signal and to interference depends on what other APs transmit. No theoretical argument in the paper shows that a local map is sufficient, and the experiments only compare averages over 200 CVX-solved points, which can hide large per-instance deviations in the min-SINR. The paper also does not compare against a centralized DNN baseline, so the observed performance cannot be attributed to the local-CSI design versus the network architecture or the teacher-student loss. The concern is therefore real and load-bearing, but it does not by itself invalidate the approach; it means the distributed/fronthaul-reduction claim needs explicit empirical support. Since the reader already rendered a CONDITIONAL verdict, my assessment does not change that verdict, hence UNCHANGED.","tokens_in":16778,"tokens_out":7021,"duration_ms":71164,"concrete_test":"Train a centralized U-net student whose input concatenates the full CSI of all L APs (i.e., stack I_l^(CAE) for l=1..L) on the same Pos-1 dataset and student loss, keeping all hyperparameters fixed. Evaluate both the centralized and distributed students on a held-out set of 1000 points, comparing average and per-sample min-SINR and SSNR against the CVX benchmark. If the centralized student matches CVX while the distributed student does not, the local-CSI premise is the bottleneck; if both match equally, the concern is resolved. Also repeat at L=4 and L=8 to test whether the gap widens with the number of APs.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central contribution is a distributed architecture in which the l-th DNN maps only the l-th AP's CSI to that AP's beamformers (Section 5.1, Eqs. (13)-(14)). But the objective being optimized is the minimum SINR over users, and Eq. (2) defines SINR_n in terms of the stacked channel h_n and stacked beamformer w_n, both spanning all APs. The useful-signal term h_n^H w_n = sum_l h_ln^H w_ln and the interference terms are sums over APs, so the optimal w_ln for a given channel realization depends on the channels and beamformers of all other APs. A function of h_ln alone cannot reproduce the centralized optimum in general; it can only hope to match it on average if the training distribution has special structure. The paper reports only averaged SSNR/SINR curves over 200 benchmark points and does not report per-sample gaps, worst-case gaps, or a comparison against a centralized DNN that receives full CSI. Consequently, the key claim that decentralized deployment with reduced fronthaul load achieves near-CVX performance rests on an unquantified approximation. This is not an internal inconsistency, but it is the most load-bearing unvalidated premise in the paper.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an unsupervised, distributed, teacher-student deep learning method for joint communication and sensing beamforming in cell-free ISAC systems. It trains one DNN per AP, where the l-th DNN maps only the l-th AP's local CSI to that AP's beamforming vectors, and trains two biased teacher models (SSNR-only and SINR-only) whose average scores are then used to normalize the student loss. The student is deployed per AP at inference time. Experiments with 1D-CNN, CAE, and U-net architectures compare average SSNR/SINR curves against a CVX-based benchmark from [28] and report a runtime comparison claiming at least three orders of magnitude speedup.","tokens_in":17014,"tokens_out":4530,"duration_ms":45955,"significance":"If the central claim holds, the work is practically significant: it offers a path to real-time, decentralized beamforming for cell-free ISAC without relying on a central processor or heavy fronthaul signaling. The unsupervised training avoids generating expensive labeled datasets, and the distributed inference structure is a genuine architectural contribution. The paper also tests multiple CNN architectures and includes a runtime comparison against a conventional solver, which is a strength. However, the significance is conditional because the evaluation does not yet support the headline quantitative claims: there is no independent test set, no statistical variability reporting, and the core distributed-CSI premise is not validated against a centralized baseline or worst-case metrics.","major_comments":[{"comment":"The distributed architecture uses the l-th DNN input as only the l-th AP's local CSI (Eqs. (13)-(14)), while the SINR objective in Eq. (2) couples all APs through the stacked channels h_n and beamformers w_lq. In general, the optimal beamformer at one AP depends on the channels and beamformers of all other APs, so a function of local CSI alone cannot reproduce the centralized optimum. The paper does not quantify the resulting gap: it never compares against a centralized DNN that receives full CSI, and it reports only average SSNR/SINR curves over 200 benchmark points without per-sample or worst-case results. Since the distributed deployment and fronthaul-load reduction claims rest on this approximation, the authors should add a focused numerical study of the local-CSI gap, including per-sample SINR distributions and, ideally, a centralized-DNN baseline.","section":"Section 5.1, Eqs. (13)-(14) and Eq. (2)"},{"comment":"The evaluation protocol is not sufficient to support the close-to-CVX performance claim. The CVX benchmark is averaged over only 200 randomly selected points with no variance indicated, and all DL training/validation curves appear to be single runs without error bars or multiple seeds. Moreover, the final reported student performance is selected post hoc from the validation curves using a 94% SINR threshold (Section 8.3), and no separate test set is held out. This makes the reported numbers vulnerable to optimistic selection bias and prevents an assessment of statistical significance. The authors should add error bars over multiple training runs, report performance on an independent test set, and state the model-selection rule before evaluation.","section":"Section 8, especially Figs. 7-10, 13-16"},{"comment":"The student loss in Eq. (11) is claimed to be bounded in [-1,0] because the denominators are said to be the 'maximum possible' scores g1(max) and g2(max). However, these values are implemented as sample means of the teacher scores over the training set, not true maxima. A student can exceed the teacher's mean, in which case g1/E[g1(max)] > 1 and the loss can fall below -1. The interpretation of the teacher models as providing strict upper bounds on student performance is therefore not mathematically valid as stated. Either the denominators should be actual maxima (with appropriate robustness handling) or the boundedness and 'maximum attainable' claims should be revised.","section":"Section 4, Eq. (11) and Section 5.2"},{"comment":"The abstract's headline 'at least three orders of magnitude' speedup is based on a runtime comparison shown in Fig. 16b, but the figure reports average runtime without confidence intervals, and the text does not specify the number of repeated trials or the variance across runs. A strong quantitative claim of this type needs a more careful timing methodology, including warm-up effects, per-sample runtime distributions, and hardware/software version details. The fairness of comparing distributed DL inference (single AP's DNN) against a centralized solver should also be discussed explicitly, although the maximum-of-two-APs choice in the paper is a reasonable start.","section":"Section 8.4 and Fig. 16b"}],"minor_comments":[{"comment":"The sentence beginning 'To address the results shown by Figure 15, we the notation, F n' is missing a verb ('introduce'), and later in the same section 'ahchieved' is a typo for 'achieved'.","section":"Section 8.3"},{"comment":"The notation block defines I_N as 'the identity vector of a size N x N', which should be 'identity matrix'.","section":"Notation and Section 2"},{"comment":"The discussion of CAE overfitting is clear, but the figures would benefit from explicit markers or legends distinguishing training and validation curves, since several curves are close together and hard to read in grayscale.","section":"Section 8.1 and Fig. 8b"},{"comment":"The lambda update rule in Eq. (12) has no convergence analysis or theoretical justification; a brief explanation of why this update should balance the two gaps would improve the presentation, even if the empirical behavior is the main evidence.","section":"Section 5.2 and Eq. (12)"},{"comment":"Some cited works are arXiv preprints rather than peer-reviewed publications; the authors should update references where stable versions exist, and should double-check the novelty statement in the Introduction in light of the recent cell-free ISAC literature.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper's core idea is interesting and the distributed teacher-student approach is worth exploring, but the evaluation currently overclaims: the local-CSI approximation is not validated, the performance numbers lack statistical support, and the model is selected on validation data without a test set. These issues are fixable within the scope of a revision, so I do not recommend rejection, but the claims in the abstract and conclusions should be scaled back until the missing evidence is supplied. I would also encourage the editor to consider whether the novelty claim about being the first DL solution for cell-free ISAC beamforming is accurate, given the nearby related work cited in the paper itself."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know this paper before you cite it: it proposes the first unsupervised teacher-student distributed beamforming scheme for cell-free ISAC, and the central claim, near-CVX performance with a runtime speedup of three orders of magnitude, is plausible but not statistically nailed down. The biggest unaddressed issue is structural: each AP's DNN sees only its own CSI (Eqs. 13-14), while the SINR objective couples all APs through the stacked channels in Eq. (2). The optimal beamformer at one AP depends in general on the other APs' channels, and the paper never quantifies the loss from this local-CSI approximation. A centralized DNN baseline or a per-sample gap analysis would settle whether the approximation actually hurts. Until then, the decentralized-fronthaul advantage rests on an unvalidated premise.\n\nWhat the paper does well: the teacher-student loss is a neat way to avoid manually tuning the sensing-communication balance; they test three CNN architectures, two position distributions, and varying user counts; they are honest about overfitting and describe their model-selection rule explicitly; the runtime comparison correctly counts the distributed case as the max of per-AP inference times. Those are real contributions.\n\nThe soft spots beyond the local-CSI issue: the benchmark is only 200 points, with no variance or confidence intervals. The DL curves are single runs without error bars across seeds. Model selection is performed on the validation set, and the final comparison uses the same validation-derived points, so there is no independent test set. That makes the reported 'close to CVX' reading noisier than the curves suggest. The paper also has minor writing issues (e.g., 'identity vector of size N×N', 'are are detailed') that a careful copyedit would fix.\n\nNone of this makes the approach implausible. The method is coherent and the results are directionally believable. But the headline claims are stronger than the evidence. This is a solid first paper for a subfield, and it deserves referee time, but the authors should be asked to provide: (1) a quantification of the local-CSI gap against a centralized DNN or at least per-sample differences, (2) results over multiple seeds with error bars and a true test set, and (3) code/data if possible. If those are added, this could become a citable baseline for distributed learning in cell-free ISAC. My recommendation: send to peer review, with those requests as major revisions.","headline":"Plausible distributed-DL scheme for cell-free ISAC, but the local-CSI approximation is unquantified and the evaluation lacks statistical rigor; still deserves a serious referee.","tokens_in":17584,"tokens_out":3139,"would_cite":true,"duration_ms":29399,"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":"An unsupervised, distributed teacher–student deep learning scheme designs cell-free ISAC beamformers whose quality is close to the convex-optimization baseline while cutting computation by at least three orders of magnitude.","keywords":["cell-free massive MIMO","integrated sensing and communication","beamforming","unsupervised learning","teacher-student training","distributed deep learning","sensing SNR","signal-to-interference-plus-noise ratio"],"falsifier":"Run the identical teacher–student procedure with more access points (for example $L=4$ and $L=8$): if the student's minimum-SINR gap to the centralized convex baseline grows steadily as $L$ increases, the claim that local-CSI-based distributed networks match the centralized solution is falsified.","tokens_in":16515,"feed_emoji":"📡","tokens_out":6373,"duration_ms":57975,"temperature":0.7,"pith_summary":"The paper tries to show that the joint sensing and communication beamforming problem of cell-free ISAC systems can be solved in real time by a distributed neural network, rather than by a centralized iterative optimizer. The proposed scheme trains one small DNN per access point; each network sees only its own AP's channel state information and outputs that AP's normalized beams for all users and the sensing target. To avoid manually balancing the sensing-versus-communication tradeoff, two teacher models are trained to extremes, one maximizing sensing SNR and the other minimum SINR, and their average scores are used to normalize the loss of a student model whose balance parameter adapts during training. The experiments with a U-net architecture report student performance close to the state-of-the-art convex solver on line-of-sight channels, with the distributed inference run time lower by at least three orders of magnitude.","feed_headline":"Distributed neural nets craft ISAC beamformers 1,000x faster","feed_subtitle":"A teacher-student scheme matches the near-optimal solver's quality while letting each access point beamform from local CSI.","key_machinery":"The load-bearing mechanism is the distributed teacher–student training unit. Two ensembles of per-AP DNNs are first trained with the teacher loss $L_t^{(\\beta)} = -[(1-\\beta)g_1 + \\beta g_2]$, one with $\\beta=0$ (sensing-only) and one with $\\beta=1$ (communication-only); their mean scores over the training set supply the reference maxima $\\hat{g}_1^{(\\max)}$ and $\\hat{g}_2^{(\\max)}$. A third identical ensemble is trained with the student loss $L_s^{(\\lambda)} = -\\left[(1-\\lambda)\\frac{g_1}{\\hat{g}_1^{(\\max)}} + \\lambda\\frac{g_2}{\\hat{g}_2^{(\\max)}}\\right]$, where $\\lambda$ is updated each mini-batch from the normalized gaps $G_1$ and $G_2$, so the sensing–communication balance is steered automatically rather than set by a fixed hyperparameter. Each DNN outputs its AP's beam set from that AP's local CSI, and the outputs are normalized to the AP power budget before SSNR and SINR are evaluated.","core_discovery":"The central claim is that a teacher–student unsupervised training procedure can produce per-access-point beamformers for cell-free ISAC that are near the quality of a centralized joint SSNR/SINR optimization, while requiring each AP to use only its local CSI at deployment. The method's contribution is that it makes the problem unsupervised and distributed: no labeled beamformer dataset is needed, the DNNs cooperate only through the loss during offline training, and after training each AP can generate its own beams independently. Reported results show the student model matching or approaching the convex baseline's average SINR at two and five users, with a slight shortfall at eight users, and an inference time around a thousandth of the baseline's.","pith_inferences":["Editorial inference: the strongest untested premise is that local CSI at each AP is enough; since the SINR of every user couples all APs, the interesting stress test is to increase the number of APs and observe whether the student-to-baseline gap widens.","Editorial inference: because the teacher scores bound what the student can reach, the method's ceiling is set by architecture capacity; the F5-with-80k-points experiment already hints that bigger filters and more data are the levers, and one could map the ceiling as a function of model size.","Editorial inference: the evaluation is restricted to line-of-sight channels and fixed AP/user position distributions; transferring to multipath or moving targets would likely require retraining or an input representation that is less geometry-specific.","Editorial inference: the 'at least three orders of magnitude' claim is an implementation-level runtime comparison on a specific CPU/GPU pairing, not an asymptotic statement, so sharper evidence would come from profiling on the actual edge hardware where the AP DNNs would run."],"forward_implications":["After offline training, each AP can compute its own beamformers from local CSI alone, so real-time deployment no longer requires the CPU to solve the optimization or send beam vectors over fronthaul for each channel realization.","Because training is unsupervised, generating a dataset does not require repeatedly solving the non-convex optimization to create labels; only channel samples and the SSNR/SINR formulas are needed.","The adaptive $\\lambda$ rule removes the need for a costly grid search over the tradeoff parameter, and the teacher-provided maxima give an architecture-specific upper bound on what the student can achieve.","In the reported line-of-sight settings the U-net student stayed close to the convex baseline as the user count varied, and the measured inference time was roughly three orders of magnitude shorter than the two-step bisection plus SDP baseline."],"supporting_citations":[{"why":"defines the cell-free ISAC joint beamforming problem and provides the convex-optimization baseline the paper compares against","marker":"[28]"},{"why":"supplies the U-net architecture that the paper adapts and that yields the best student performance","marker":"[49]"},{"why":"motivates the distributed DNN approach that avoids fronthaul information collection in cell-free networks","marker":"[36]"},{"why":"provides the decentralized unsupervised training paradigm for hybrid beamforming in cell-free systems","marker":"[38]"},{"why":"shows an unsupervised cell-free power-control approach whose centralized input motivates the authors' distributed alternative","marker":"[37]"},{"why":"the prior deep-reinforcement-learning cell-free ISAC power allocation work that this paper contrasts as needing discrete actions and not solving general beamforming","marker":"[9]"},{"why":"inspires the teacher-student training idea used to set the sensing-communication balance","marker":"[44]"}],"fun_headline_variants":["Teacher-student nets craft ISAC beams 1000x faster","Unsupervised DNN beamforming for cell-free ISAC 1000x faster","Local CSI neural net beamforming matches solver at 1000x speed","Distributed unsupervised learning beamforms ISAC at 1000x"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that each AP's neural network can produce near-optimal beamformers using only that AP's local channel information, even though the SINR expression couples every AP's channels and beams through each user's receiver.","fun_headline_variants_meta":{"raw":{"variants":["Teacher-student nets craft ISAC beams 1000x faster","Unsupervised DNN beamforming for cell-free ISAC 1000x faster","Local CSI neural net beamforming matches solver at 1000x speed","Distributed unsupervised learning beamforms ISAC at 1000x"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000924,"raw_usage":{"total_tokens":3963,"prompt_tokens":950,"completion_tokens":3013,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":566,"completion_tokens_details":{"reasoning_tokens":2933}},"tokens_in":566,"tokens_out":3013,"duration_ms":21214,"temperature":1.0,"reasoning_tokens":2933,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T04:58:31.918974+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the identical teacher–student procedure with more access points (for example $L=4$ and $L=8$): if the student's minimum-SINR gap to the centralized convex baseline grows steadily as $L$ increases, the claim that local-CSI-based distributed networks match the centralized solution is falsified.","supporting_citations":[{"cited_title":"Ronneberger, P","cited_arxiv_id":null,"evidence_quote":"supplies the U-net architecture that the paper adapts and that yields the best student performance"},{"cited_title":"Zaher, ¨O","cited_arxiv_id":null,"evidence_quote":"motivates the distributed DNN approach that avoids fronthaul information collection in cell-free networks"},{"cited_title":"Hojatian, J","cited_arxiv_id":null,"evidence_quote":"provides the decentralized unsupervised training paradigm for hybrid beamforming in cell-free systems"},{"cited_title":"Nikbakht, A","cited_arxiv_id":null,"evidence_quote":"shows an unsupervised cell-free power-control approach whose centralized input motivates the authors' distributed alternative"}],"review_version":1}