{"id":"bb967015-ed24-4252-acb1-fe58b2cded08","arxiv_id":"2508.07009","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A cascaded Transformer neural network predicts ergodic spectral efficiency from user positions, enabling a low-complexity stable-matching scheduler for multi-user active-IRS systems.","lead":"This paper builds a machine-learning system that predicts wireless signal quality from user positions and past data, then uses those predictions to schedule users across multiple active reflecting surfaces. The aim is to skip costly real-time channel measurement while keeping the slowest user as fast as possible.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Near-optimality is shown only against a Gurobi upper bound that uses the same neural CKM predictions; true max-min throughput is never computed, so the central claim is not yet substantiated.","rationale":"The paper has useful contributions: the cascaded Transformer-based neural CKM, the low-complexity SM-IB scheduler, and numerical evidence that the proposed networks outperform MLP/LSTM baselines. However, the headline 'near-optimal max-min throughput' relies on comparing SM-IB with a Gurobi upper bound that is computed on the same neural-CKM surrogate objective. Section VI-C explicitly grounds the evaluation in the trained neural CKM, and no true Monte-Carlo ergodic SE is computed for the final schedules or for the upper bound. This is not a peripheral issue: max-min optimization is sensitive to systematic prediction bias at the worst-off UE, and Table I's average MAE/MRE does not bound such bias. The reader's weakest assumption concerned spatial generalization outside the training distribution; my concern is internal and more direct—the evaluation protocol does not substantiate the claim even in the tested scenario. The reader's verdict remains CONDITIONAL because a concrete true-objective check could resolve the concern, and the paper's other components are promising. If the check failed, the verdict would need to move toward REJECT or UNVERDICTED, but that is not determined yet.","tokens_in":11268,"tokens_out":10838,"duration_ms":117064,"concrete_test":"Re-evaluate the smallest benchmark (e.g., I=6, U=30) with true ergodic SEs: for the SM-IB final schedule, compute each UE's SE via Monte Carlo (50 large-scale × 500 small-scale QuaDRiGa realizations, as in §VI-A); compute the true max-min throughput. Also obtain a true optimal max-min by running Gurobi on (P2) with Monte-Carlo-computed SEs (feasible for U=30). If the true SM-IB value is within the reported gap (or ~10%) of the true optimum, the concern is resolved; if not, the near-optimality claim must be revised.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim that SM-IB achieves near-optimal max-min throughput is only validated against an upper bound computed with the same learned surrogate. Section VI-C states: 'Based on the trained neural CKM, we can further evaluate the performance of the proposed SM-IB algorithm' and the Gurobi upper bound solves (P2) using the same SE predictions (Eqs. 16-17). Thus Fig. 4's gap measures scheduling suboptimality under the neural-CKM model, not under true channel statistics. Because an optimizer can exploit even small average prediction biases—especially for max-min fairness, where the bottleneck UE's predicted SE may be systematically overestimated—the reported near-optimality can be an artifact of the surrogate. The paper never evaluates true Monte-Carlo ergodic SEs for the final schedules nor computes a true upper bound, leaving the headline claim unsupported. This is distinct from (though compounding) the reader's spatial-generalization concern: even in the tested scenario, the objective being optimized and evaluated is the surrogate's.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a scheduling framework for multi-user downlink systems with multiple active IRSs. Because instantaneous CSI acquisition is costly, the authors introduce a neural Channel Knowledge Map (CKM) composed of two cascaded Transformer-based networks: LPS-Net, which predicts per-link power statistics (CDF quantiles) from user position and system configuration, and SE-Net, which maps these statistics and the UE-to-IRS association to an ergodic spectral efficiency. The ergodic SE predictions are then used as the objective in a max-min throughput scheduling problem (P2), solved by a three-stage Stable Matching-Iterative Balancing (SM-IB) algorithm. Numerical results on QuaDRiGa-generated channels report that the neural CKM improves prediction accuracy over MLP/LSTM baselines and that SM-IB closely approaches a Gurobi-based upper bound while reducing runtime.","tokens_in":11466,"tokens_out":3195,"duration_ms":37206,"significance":"If the claims hold, the paper offers a practical way to avoid instantaneous CSI in IRS scheduling, with a modular learning architecture (predict link statistics, then compose them into SE) and a low-complexity heuristic for a difficult mixed-integer problem. The problem is timely and the proposed decomposition is sensible. The paper also provides explicit complexity expressions and a comparison against Gurobi. However, the central validation is incomplete: the reported near-optimality is measured against an upper bound computed on the same neural-network surrogate, not on true channel statistics, and the spatial generalization of the learned CKM is not tested beyond a random split of the same 10,000-sample QuaDRiGa dataset. These gaps are load-bearing for the paper's headline claims.","major_comments":[{"comment":"The 'near-optimal' claim compares SM-IB only against a Gurobi solution of (P2) in which the objective Ru is computed from SE-Net predictions (Eq. (17)), not from true ergodic spectral efficiency. The paper never evaluates the true Monte-Carlo ergodic throughput of the final schedules, nor does it compute a true upper bound. In a max-min problem, systematic overestimation of the bottleneck UE's SE can make the heuristic appear near-optimal even when actual fairness is poor. The authors should report the actual ergodic throughput of the schedules produced by SM-IB and by Gurobi using the channel model (Eq. (5)), ideally with confidence intervals, and, for small instances, compare against exhaustive search or a valid true upper bound.","section":"§VI-C, Eq. (17), Fig. 4"},{"comment":"The spatial-generalization premise is not validated. LPS-Net and SE-Net are trained on 10,000 QuaDRiGa samples and evaluated on a random 8:1:1 split of the same dataset. Since SM-IB is later applied to arbitrary UE positions and scheduling configurations, prediction bias outside the training distribution could make the algorithm optimize a systematically wrong objective. The authors should add cross-scenario validation, e.g., testing on UE positions, AIRS locations/orientations, or shadowing maps not seen in training, and report prediction error as a function of distance from the training positions. Without this, the scheduling results in Fig. 4 are only valid under the untested assumption that the learned mapping generalizes.","section":"§IV-B, §VI-B"},{"comment":"No error bars, confidence intervals, or multiple random trials are reported for the key scheduling comparison. Given the stochasticity of QuaDRiGa large-scale parameters, small-scale fading, and neural-network training, the statement that the proposed algorithm 'achieves a max-min ergodic throughput close to that of the upper bound' is not statistically supported. The authors should report means with confidence intervals over multiple independent dataset/training/scheduling runs, or at least over multiple random UE drops.","section":"§VI-C, Fig. 4, Table II"}],"minor_comments":[{"comment":"The loss function for SE-Net is denoted fδLPS(η̄, η̂), but the parameter used is δSE. This notation likely refers to the same smooth L1 loss, but should be renamed (e.g., fδSE) to avoid confusion.","section":"Eq. (20)"},{"comment":"The table header 'NetworkNpara.' is unclear; define Npara as the number of trainable parameters in the caption or main text.","section":"Table I"},{"comment":"The complexity of the Iterative Balancing algorithm is given as O(log2(1/ϵ)) without derivation. Since the throughput gap may not shrink geometrically in all cases, a brief justification or reference would strengthen the claim.","section":"§V-B"},{"comment":"The comparison of phase optimization schemes reports single-point values (3.233 versus 3.357 bps/Hz) without variance or confidence intervals; please add error bars or note if these are means over the 50 realizations.","section":"§VI-A"},{"comment":"No code or data is provided. Given the reliance on a simulated dataset and deep learning, releasing code/data or at least detailed experiment seeds would improve reproducibility.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The paper's core idea is reasonable and the experiments are extensive for a simulation study. The main concern is that the central 'near-optimal' claim is validated only against a surrogate-based upper bound, and the spatial generalization of the neural CKM is untested. Both issues can be addressed with additional experiments (true Monte-Carlo evaluation of final schedules, cross-scenario validation), so I do not recommend rejection. The lack of code/data is secondary but worth requesting."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea here is genuinely useful: replace per-scheduling CSI acquisition with a learned map from UE positions to ergodic SE, then schedule on top of that map. The cascaded LPS-Net/SE-Net design is new in this context, and the SM-IB heuristic is a sensible, low-complexity matching/balancing scheme. The prediction accuracy numbers (MRE ~8.65% for SE, quantile errors under 1 dB for link power CDFs) beat the MLP/LSTM baselines, and the running-time advantage over Gurobi is dramatic. On the strengths, this paper does a credible job of formulating a hard scheduling problem and showing that a learned surrogate plus a tailored heuristic can make it tractable.\n\nBut the central claim—near-optimal max-min throughput—is not actually substantiated. The Gurobi upper bound in Fig. 4 solves (P2) using the same trained neural CKM predictions, not true Monte Carlo ergodic SEs. So the gap shown is the scheduling suboptimality under the surrogate model, and it says nothing about how close SM-IB is to the true optimum under actual channel statistics. The paper never evaluates the final schedules with true MC ergodic SE, nor computes a true upper bound. That is a load-bearing omission. Even in the tested scenario, the optimized objective is the surrogate's, and max-min fairness is exactly the kind of objective where small prediction biases on bottleneck UEs can be exploited.\n\nOther soft spots are more minor but still matter: no error bars on the key scheduling plot, no code or data release, a single QuaDRiGa scenario, and convergence arguments for SM-IB that are heuristic (monotonic improvement, not optimality). The loss weights and hyperparameters look hand-picked rather than systematically tuned. None of this kills the proposal, but it does mean the paper's strongest claim is conditional.\n\nThe paper is written coherently, and I don't see internal contradictions. It deserves a serious referee: the problem is real, the architecture is novel enough, and the empirical evidence for prediction quality is decent. But it needs revision before publication: evaluate the final schedules with true MC ergodic SE, add error bars, release the artifacts, and temper the 'near-optimal' language until it's actually measured. I'd bring it to a reading group if the discussion is about learning-based resource allocation; I wouldn't cite it yet for the scheduling result.","headline":"Useful learned-CKM scheduling framework for multi-AIRS systems, but the 'near-optimal' headline is only proven against the same neural surrogate, so the true gap to optimality remains open.","tokens_in":12011,"tokens_out":1472,"would_cite":false,"duration_ms":16619,"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":"This paper claims that a Transformer-based channel knowledge map trained on position-tagged historical measurements can predict ergodic spectral efficiency accurately enough for a low-complexity matching scheduler to reach near-optimal max-","keywords":["neural channel knowledge map","active IRS","ergodic spectral efficiency prediction","max-min throughput scheduling","stable matching","iterative balancing","Transformer","multi-user OFDMA"],"falsifier":"Train the neural CKM on samples whose user positions lie within a small central region of a cell, then evaluate it for users spread across the full cell or a different cell layout; if predicted ergodic SE drifts from Monte Carlo averages (substantial mean-relative-error growth) at unvisited positions, the spatial-generalization premise fails. A deployment test does the same with live channel measurements: compare the CKM's predicted link-power CDFs against measured ones at positions never seen in training, and check whether the SM-IB schedule computed from them still attains the predicted max-","tokens_in":11116,"feed_emoji":"📡","tokens_out":8631,"duration_ms":87050,"temperature":0.7,"pith_summary":"This paper tries to remove the need for real-time channel state information when scheduling many users across multiple active IRSs. It builds a neural channel knowledge map that learns, from historical measurements tagged with user positions, the distribution of each link's received power and then the resulting ergodic spectral efficiency under any candidate association. On top of that predictor it runs a low-complexity scheduler, SM-IB, to maximize the minimum achievable ergodic throughput subject to each IRS serving one user per slot. The paper demonstrates in simulation that the predictor is more accurate than MLP and LSTM baselines, and that the scheduler comes close to the exact solver's upper bound in max-min throughput while running in milliseconds to a few tenths of a second. A sympathetic reader would see this as making cell-level multi-AIRS scheduling a matter of cheap position-based inference rather than expensive instantaneous CSI acquisition.","feed_headline":"Learn the map, skip the CSI: IRS scheduling near-optimal at 0.1 seconds","feed_subtitle":"Position-tagged history replaces live channel state; stable matching keeps fairness near the exact optimum at a fraction of the runtime.","key_machinery":"The central object is the neural channel knowledge map, split into two learned maps. LPS-Net is a Transformer encoder with piecewise-linear encoded scalar features and learnable target tokens; it maps a UE position plus AIRS configuration to a vector of 16-quantile CDFs for direct, cascaded, scattered, and dynamic-noise link powers. SE-Net is a Transformer encoder whose input tokens are the CDFs of the direct, serving, non-serving, and noise links for a given association, with a category position embedding, and whose output token is regressed to the ergodic SE. The scheduling machinery is SM-IB: constrained K-Means clusters UEs, Gale-Shapley stable matching forms UE-to-AIRS associations, ite","core_discovery":"The paper claims that in a single-cell downlink network with multiple active IRSs, the difficult job of scheduling users across IRSs and resource blocks can be separated from instantaneous channel state acquisition. It trains two cascaded Transformer-based networks on historical, position-tagged channel and throughput measurements: LPS-Net predicts per-link power statistics (quantized CDFs of direct, cascaded, scattered, and noise powers) from a UE position and the AIRS configuration, and SE-Net composes those link statistics, under a candidate UE-to-AIRS association, into an ergodic spectral efficiency estimate. On these predicted efficiencies, a three-stage SM-IB scheduler performs stable","pith_inferences":["Because LPS-Net outputs link-power CDFs rather than a single rate, the same trained network could support objectives beyond max-min throughput, such as outage-constrained or delay-sensitive scheduling, by re-designing only the SE-Net head; the paper does not explore this.","The reported near-optimality is demonstrated for a particular inner loop with equal transmit-power allocation and a fixed MCCM beamforming scheme; extending the same neural-CKM predictor to jointly optimize power and phase would require re-running the gap analysis, not just retraining the networks.","A natural testable extension is mobility-aware scheduling: since the map is position-conditioned, predicted SEs for predicted future locations could drive proactive association and slot assignment, sidestepping the latency of fresh CSI at every reposition.","The SE-Net architecture treats link CDFs as a variable-length token set, so the same framework could scale to more or fewer AIRSs than the 2 to 6 shown, provided the training set covers the enlarged configuration space."],"forward_implications":["Under the paper's assumptions, cell-level multi-AIRS scheduling can be run from position fixes and historical measurements, so channel-state acquisition delay no longer makes scheduling decisions obsolete.","The cascaded LPS-Net and SE-Net decomposition means the same link-power statistics can be reused for many candidate associations, cutting the cost of evaluating schedules from channel simulation to milliseconds.","SM-IB's hierarchy gives max-min throughput close to the exact solver's upper bound with runtime that stays around a tenth of a second even at 210 users, while the exact solver's runtime grows to tens of minutes.","Since the predictor is trained on link-power CDFs rather than raw channels, the same architecture can adapt to other AIRS phase optimization schemes without retraining the whole stack.","Max-min fairness becomes tractable at cell scale: the paper's numbers suggest that 6 AIRSs and 210 users can be scheduled in about a third of a second, enabling dynamic re-scheduling as users move."],"supporting_citations":[{"why":"Defines the CKM concept of mapping positions to channel knowledge, motivating the whole approach of bypassing real-time full CSI acquisition.","marker":"[10]"},{"why":"Supplies the assumption that IRS-related channel power statistics can be obtained from statistical measurements, which LPS-Net learns to predict.","marker":"[5]"},{"why":"Provides the Transformer encoder architecture used as the core building block of LPS-Net and SE-Net.","marker":"[13]"},{"why":"Supplies the low-complexity MCCM beamforming scheme whose phase coefficients define the cascaded channels and hence the SE labels.","marker":"[16]"},{"why":"Provides the Gale-Shapley stable matching procedure used for UE-to-AIRS association in SM-IB.","marker":"[18]"},{"why":"Introduces the active IRS model with amplification and dynamic noise, which grounds the SNR and ergodic SE expressions.","marker":"[3]"},{"why":"Provides the capacity-characterization upper bound that justifies the choice of the MCCM phase optimization scheme in the numerical study.","marker":"[15]"},{"why":"Supplies the piecewise-linear encoding used to map scalar configuration features into Transformer input tokens.","marker":"[17]"}],"fun_headline_variants":["Neural map replaces live CSI for near-optimal IRS scheduling","Position history predicts IRS links, scheduling near-optimal fast","Cascaded Transformers predict spectral efficiency, stable matching schedules","IRS scheduling from historical CSI: near-optimal at low complexity","Smart map + stable matching: IRS scheduling near-optimal speed"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The load-bearing premise is spatial generalization: a regression model trained on one dataset of simulated channel samples at known positions is assumed to predict the ergodic spectral efficiency for any user position and system configuration the optimizer will meet; if that learned map is biased outside its training distribution, the scheduler optimizes a wrong objective and the near-optimality result collapses.","fun_headline_variants_meta":{"raw":{"variants":["Neural map replaces live CSI for near-optimal IRS scheduling","Position history predicts IRS links, scheduling near-optimal fast","Cascaded Transformers predict spectral efficiency, stable matching schedules","IRS scheduling from historical CSI: near-optimal at low complexity","Smart map + stable matching: IRS scheduling near-optimal speed"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000191,"raw_usage":{"total_tokens":1177,"prompt_tokens":738,"completion_tokens":439,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":482,"completion_tokens_details":{"reasoning_tokens":354}},"tokens_in":482,"tokens_out":439,"duration_ms":4789,"temperature":1.0,"reasoning_tokens":354,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T22:23:36.095902+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the neural CKM on samples whose user positions lie within a small central region of a cell, then evaluate it for users spread across the full cell or a different cell layout; if predicted ergodic SE drifts from Monte Carlo averages (substantial mean-relative-error growth) at unvisited positions, the spatial-generalization premise fails. A deployment test does the same with live channel measurements: compare the CKM's predicted link-power CDFs against measured ones at positions never seen in training, and check whether the SM-IB schedule computed from them still attains the predicted max-","supporting_citations":[{"cited_title":"A tutorial on environment-aware communications via channel knowledge map for 6g,","cited_arxiv_id":null,"evidence_quote":"Defines the CKM concept of mapping positions to channel knowledge, motivating the whole approach of bypassing real-time full CSI acquisition."},{"cited_title":"Multi-pair D2D communications aided by an active RIS over spatially correlated channels with phase noise,","cited_arxiv_id":null,"evidence_quote":"Supplies the assumption that IRS-related channel power statistics can be obtained from statistical measurements, which LPS-Net learns to predict."},{"cited_title":"An image is worth 16x16 words: Transformers for image recognition at scale,","cited_arxiv_id":null,"evidence_quote":"Provides the Transformer encoder architecture used as the core building block of LPS-Net and SE-Net."},{"cited_title":"Beam-squint mitigating in reconfig- urable intelligent surface aided wideband mmwave communications,","cited_arxiv_id":null,"evidence_quote":"Supplies the low-complexity MCCM beamforming scheme whose phase coefficients define the cascaded channels and hence the SE labels."},{"cited_title":"Matching theory: Applications in wireless communica- tions,","cited_arxiv_id":null,"evidence_quote":"Provides the Gale-Shapley stable matching procedure used for UE-to-AIRS association in SM-IB."},{"cited_title":"Active RIS vs. passive RIS: Which will prevail in 6G?","cited_arxiv_id":null,"evidence_quote":"Introduces the active IRS model with amplification and dynamic noise, which grounds the SNR and ergodic SE expressions."},{"cited_title":"Capacity characterization for intelligent reflect- ing surface aided MIMO communication,","cited_arxiv_id":null,"evidence_quote":"Provides the capacity-characterization upper bound that justifies the choice of the MCCM phase optimization scheme in the numerical study."},{"cited_title":"On embeddings for numerical features in tabular deep learning,","cited_arxiv_id":null,"evidence_quote":"Supplies the piecewise-linear encoding used to map scalar configuration features into Transformer input tokens."}],"review_version":1}