{"id":"66facd1d-798e-445a-b85e-f5f21b52d203","arxiv_id":"2505.23198","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A deep learning Wi-Fi channel feedback system with trainable vector quantization, angle-difference reporting, and access-point refinement outperforms the standard 802.11 feedback in simulations.","lead":"This paper proposes a deep learning system that compresses Wi-Fi channel information into small codebooks and uses past channel reports to improve reconstructions at the access point. The approach is tested on synthetic and real Wi-Fi channels and is reported to improve accuracy and net throughput over the standard 802.11 feedback method, which matters because feedback overhead is a bottleneck in next-generation Wi-Fi.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table VI's net-throughput comparison appears to assume one 2000-byte packet per feedback cycle, inconsistent with the paper's 40 Hz feedback setting; the large throughput gains over 802.11 may be an artifact of this timing assumption.","rationale":"I read the paper as a careful empirical study of VQ-based angle feedback for Wi-Fi, and the NMSE comparisons in Figs. 11-13 and Table III are internally reasonable. My concern is not with the learned-codebook machinery but with the net-throughput analysis attached to the paper's headline claim. The net throughput numbers in Table VI are only meaningful if the overhead-to-data timing is specified correctly. The paper specifies a 40 Hz feedback frequency in Table I, and Eq. (28) claims to amortize overhead over Ndata. Yet the numerical ratios imply Ndata is essentially one 2000-byte packet, meaning feedback is modeled as occurring roughly every 100 microseconds rather than every 25 milliseconds. Under the stated periodic-feedback scenario, the overhead difference between a 577-bit report and a 9984-bit report is small relative to the 25 ms interval, so the net throughput should be close to the gross rate for all schemes. I therefore disagree with the reader's choice of perfect channel estimation as the weakest assumption: that is a real limitation, but the throughput timing inconsistency is more directly load-bearing for the strongest claim and is checkable from the paper's own parameters. A corrected throughput simulation could still show modest gains, but the current 3.3x advantage is not credible. This does not change the conditional verdict, because the NMSE results and the methods still merit scrutiny and revision; it does add a specific, mandatory condition: rerun the throughput comparison with a realistic feedback interval and state the packet-count assumption. No code or error bars are provided, which further supports keeping the verdict conditional.","tokens_in":20438,"tokens_out":12337,"duration_ms":140108,"concrete_test":"Recompute Table VI with a 40 Hz feedback interval: set Ndata = 25 ms minus Noverhead and allow multiple 2000-byte packets between feedback frames, or add an explicit packet-count parameter to Eq. (28). If the proposed 577-bit scheme's net throughput stays within a few percent of 802.11 T0/T1 under this amortized overhead, the reported throughput advantage is an artifact of single-packet-per-feedback timing.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim includes net throughput gains over the 802.11 standard, but the throughput model in Sec. V-D is not tied to the simulated 40 Hz feedback frequency (Table I). Equation (28) defines Ndata as the data-transmission duration, yet the implementation appears to set Ndata to the duration of a single 2000-byte packet. The implied ratios in Table VI confirm this: standard T0 has 19.37/162.5 = 0.119, so Noverhead is about 7.4 times Ndata; with a 2000-byte packet at 162.5 Mb/s (about 98.5 microseconds), Noverhead is about 0.73 ms, which is a plausible CBR plus NDPA/NDP/ACK overhead but equivalent to feedback once per packet, i.e., thousands of times per second. Under the stated 40 Hz periodic feedback, overhead is paid once per 25 ms, Ndata would be roughly 24 to 25 ms, and net throughput for standard T0 would be about 158 to 160 Mb/s instead of 19.37 Mb/s, while the proposed 577-bit scheme would be about 161 to 162 Mb/s. The advertised 3.3x net-throughput gain collapses to a few percent. The net-throughput component of the central claim is therefore not supported by the simulation as currently described.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a deep-learning-based CSI feedback framework for Wi-Fi systems. The STA compresses the IEEE 802.11 Givens-rotation angle parameters through an encoder and a trainable vector-quantization module, and the AP reconstructs them with a decoder. To exploit temporal correlation, the authors introduce an angle-difference feedback strategy (with a preprocessing step for angle periodicity and two VQ designs, parallel and unified) and a CSI refinement module, including a recursive refinement strategy, that operates on the AP side. The methods are evaluated with the DeepMIMO and Wi-MIR datasets in terms of NMSE, complexity, and net throughput, and are compared with standard 802.11 feedback, two prior DL-based Wi-Fi CSI feedback methods, and two prior refinement architectures.","tokens_in":20724,"tokens_out":2541,"duration_ms":27741,"significance":"If the reported results hold, the paper makes a useful contribution to the relatively underexplored area of DL-based CSI feedback for Wi-Fi systems, where prior work has not combined trainable vector quantization with temporal-correlation exploitation. The empirical methodology is a strength: two datasets are used, including a real-measured dataset; comparisons include standard and prior DL baselines; and complexity in parameters, FLOPs, and runtime is reported. The proposed angle-difference preprocessing and recursive refinement are concrete, falsifiable design choices. The central NMSE claims appear internally consistent with the described training procedures, but the net-throughput claim, the absence of statistical uncertainty, and the perfect-channel-estimation assumption require attention before the paper's conclusions can be fully accepted.","major_comments":[{"comment":"The net-throughput comparison is not tied to the stated feedback period. Table I sets the feedback frequency to 40 Hz, but Eq. (28) defines Ndata as the data-transmission duration and the implementation appears to set Ndata to the duration of a single 2000-byte packet. The implied overhead ratio in Table VI (standard T0: 19.37/162.5 = 0.119, so Noverhead is about 7.4 times Ndata) corresponds to feedback being transmitted roughly once per packet, i.e., thousands of times per second, not once per 25 ms. With the stated 40 Hz periodicity, the standard T0 net throughput would be about 158-160 Mb/s rather than 19.37 Mb/s, and the proposed 577-bit scheme would be about 161-162 Mb/s. The advertised 3.3x net-throughput gain over the standard therefore collapses to a few percent under the paper's own feedback-frequency setting. The net-throughput component of the central claim is not supported as currently described and must be corrected or removed.","section":"Sec. V-D, Eq. (28), Table VI"},{"comment":"The paper assumes perfect channel estimation at the STA. The entire pipeline, including SVD, angle extraction, threshold decisions in Eq. (13)-(14), and codebook training, is built on noiseless channels. In a real Wi-Fi system, channel estimation errors will alter the distribution of the angle parameters and may degrade the learned codebooks and the sparse-difference assumption. Since the central claim is about practical Wi-Fi systems, the authors should either quantify the sensitivity of the reported NMSE and throughput gains to estimation errors or clearly delimit the claims to the perfect-CSI setting.","section":"Sec. II-A"},{"comment":"All reported NMSE and throughput values are point estimates. No error bars, confidence intervals, or multiple-seed results are provided, despite the stochastic nature of neural-network training and the small performance differences at higher feedback overheads (e.g., Table VI shows a 1-2% net-throughput difference between some schemes). Without this information, it is difficult to assess whether the claimed gains, especially the smaller ones, are statistically reliable. The authors should report the variance across at least a few independent training runs or justify why the reported differences are robust.","section":"Sec. V-A and Figs. 11-13, Tables III, VI"},{"comment":"The sparsity thresholds (mu_th, N_th) are set separately for the two datasets with no stated criterion or sensitivity analysis. Because the indicator It in Eq. (14) determines whether the angle-difference feedback is used at all, the reported gains of the angle-difference strategy depend directly on these hand-tuned values. The authors should provide a sensitivity study or a principled selection rule to show that the conclusions do not hinge on the particular threshold choices.","section":"Sec. V-A and Sec. III-C"}],"minor_comments":[{"comment":"The word \"inital\" in the first sentence should be \"initial.\"","section":"Sec. III-A"},{"comment":"In the sentence after Eq. (13), \"where mu_th is a threshold a threshold below which values are considered zero\" contains a duplicated phrase.","section":"Sec. III-C"},{"comment":"The phrase \"For the CSI feedback, which is corresponded to CBR\" should read \"which corresponds to CBR.\"","section":"Sec. V-D"},{"comment":"The hyperparameters D=16, B=8, and beta=0.25 are stated once; it would be helpful to also state the codebook size and the number of sub-vectors explicitly for reproducing the results.","section":"Sec. V-A"},{"comment":"The table structure repeats the scheme names in each sub-block; a cleaner layout would improve readability, though this is purely presentational.","section":"Table VI"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for a signal processing or wireless communications venue, and the NMSE-based contributions are technically plausible. My main concern is that the net-throughput claim, which is part of the stated contributions, is not supported by the current simulation setup. This is fixable in revision, so I do not recommend rejection. I would also encourage the editor to ensure that the authors provide training/seed details or code, since the empirical claims currently lack uncertainty quantification."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Candid take: the real contribution here is the angle-difference feedback plus AP-side refinement, and the NMSE results support it. The net-throughput part is not credible as presented, and the stress-test concern holds up on reading.\n\nIn Sec. V-D, Eq. (28) defines Ndata as the data-transmission duration, but the implementation clearly uses one 2000-byte packet per feedback cycle. For standard T0, net/gross = 19.37/162.5 ≈ 0.119, which implies Ndata ≈ 98.5 µs and Noverhead ≈ 0.73 ms. That is per-packet overhead, not overhead per feedback interval. At the paper's stated 40 Hz feedback, the feedback overhead is paid once per 25 ms, so the standard's net throughput would be roughly 160 Mb/s and the proposed 577-bit scheme around 162 Mb/s. The advertised 3.3x gain collapses to a few percent. Table VI is therefore misleading.\n\nThe NMSE work, though, is a legitimate advance. Trainable vector quantization for the IEEE 802.11 angle parameterization, the parallel and unified VQ designs to handle error propagation, and the recursive refinement using current plus previous feedback are genuinely new in this Wi-Fi CSI feedback subfield. The experiments are reasonable: two datasets including a real one, comparisons against both standard feedback and prior DL baselines, and a complexity table. The lack of error bars and multiple seeds is a minor concern, and perfect-CSI is a standard assumption for this line of work, though it should be listed as a limitation. The per-dataset thresholds (mu_th, N_th) look hand-tuned, but that is common in this literature.\n\nSo: the central NMSE claims are plausible and worth publishing once the throughput analysis is fixed. The paper deserves a serious referee. I would send it to review, but instruct the reviewers to require a corrected net-throughput model that respects the stated feedback period. I would bring it to the reading group and would likely cite the NMSE design if I work on CSI feedback.","headline":"Solid NMSE contribution to DL-based CSI feedback for Wi-Fi, but the net-throughput table is built on a per-packet overhead assumption that collapses the advertised 3x gain.","tokens_in":21299,"tokens_out":2229,"would_cite":true,"duration_ms":23942,"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 learned angle codec with temporal differencing beats IEEE 802.11 CSI feedback in simulations.","keywords":["CSI feedback","deep learning","vector quantization","angle-difference feedback","temporal correlation","IEEE 802.11","beamforming","channel state information"],"falsifier":"Retrain and evaluate the same framework on the same DeepMIMO and Wi-MIR datasets but with channel estimates corrupted by the estimator used in NDP-based 802.11 acquisition, such as least-squares with the same pilot structure and receiver noise; if, at equal feedback overhead, the NMSE gap over the standard 802.11 quantizer disappears or the angle-difference mode rarely triggers, the claim of consistent gains is falsified.","tokens_in":20206,"feed_emoji":"📶","tokens_out":9835,"duration_ms":87333,"temperature":0.7,"pith_summary":"The paper proposes replacing the IEEE 802.11 CSI feedback procedure—uniform quantization of Givens beamforming angles—with a learned autoencoder that compresses these angles into a finite-bit representation via a trainable vector-quantization codebook, jointly trained end-to-end. On top of this baseline, it adds an angle-difference feedback strategy: when the wrapped difference between consecutive angle frames is sparse, the STA sends the difference rather than the full angles, using either a parallel or a unified VQ design to prevent reconstruction errors from accumulating across time. It also adds an AP-side refinement module, built from a video-prediction network, that reconstructs the current angles using both current and past feedback. The combined framework is shown to outperform the standard IEEE 802.11 feedback and earlier DL-based Wi-Fi feedback methods in NMSE across synthetic and real channel datasets, and to raise net throughput in the simulated 802.11be setup. The relevance is practical: next-generation Wi-Fi needs accurate downlink channel knowledge at low feedback overhead, and the paper's central claim is that temporal correlation makes that feedback nearly free.","feed_headline":"Learned codec cuts Wi-Fi CSI feedback, lifts throughput","feed_subtitle":"Trainable vector-quantized autoencoder plus angle-difference feedback beats the 802.11 standard in simulated Wi-Fi.","key_machinery":"The load-bearing identity is the angle-parameter matrix $\\Phi_t = [\\{\\phi_{j,i}\\}, \\{\\psi_{l,i}\\}]$ per subcarrier, extracted from the SVD of the channel via Givens rotations exactly as in the IEEE 802.11 standard; using these angles as the encoder input (instead of the full beamforming matrix $\\mathbf{V}$) shrinks the input and preserves the orthogonality structure without output constraints. Around this input, the paper trains an encoder–decoder pair with a product vector-quantization codebook (the VQ-VAE loss of Eq. (11)), a preprocessing map $h(\\cdot)$ that wraps differences into $[-\\pi, \\pi)$ to handle angle periodicity, a sparsity trigger comparing the count of nonzero wrapped differences to thresholds $(\\mu_{\\mathrm{th}}, N_{\\mathrm{th}})$, and two feedback architectures—parallel VQ, a two-stage codebook that spends $B_r$ bits on the previous latent residual, and unified VQ, which sums the previous residual latent into the current encoder output before a single quantizer, eliminating the $B_r$ hyperparameter. The refinement module is a SimVP-style fully convolutional predictor that maps the last $T$ reconstructed angle frames to a refined current frame, trained after freezing the feedback network and then recursively.","core_discovery":"The paper's claim is that CSI feedback for Wi-Fi can be treated as a learned compression problem on the Givens-rotation angle parameters $(\\phi, \\psi)$ that the standard already extracts, and that a codec trained with a product vector-quantization module in the latent space is strictly more efficient than the standard's uniform scalar quantization at equal feedback cost. The second claim is that because consecutive angle frames are temporally correlated, transmitting the preprocessed wrapped difference $h(\\Phi_t - \\Phi_{t-1})$ instead of the full angles, gated by a sparsity threshold, yields substantial additional gains, provided error propagation is suppressed by either the parallel VQ (which quantizes the previous residual with a reserved bit budget) or the unified VQ (which adds the previous residual latent to the current encoder output before the single codebook). The third claim is that an AP-side refinement module, using the SimVP CNN on the last $T$ reconstructed frames, further improves reconstruction NMSE, and a recursive variant that feeds refined outputs back into future refinement improves it further; the full system is reported to outperform standard IEEE 802.11 feedback (T0/T1) in both NMSE and net throughput on the DeepMIMO and Wi-MIR datasets.","pith_inferences":["Since the sparsity trigger determines when the angle-difference mode is used, the framework's gains should shrink as feedback rate falls or STA speed rises; an implicit bound is the point where the trigger never fires, making the system reduce to initial feedback, and the speed sweep suggests that knee lies above 2.8 m/s.","The paper freezes the feedback network before training the refinement module, so the AP can switch refinement on and off without retraining; jointly training feedback and refinement with a two-phase schedule might yield larger gains but risks degrading the no-refinement operating point.","The VQ codebook indices are transmitted as fixed-length codewords; applying entropy coding to them, as the paper mentions as future work, would reduce feedback bits further without changing reconstruction since the decoder only sees indices.","All results assume perfect channel estimation at the STA; with real NDP-based estimation noise, the angle differences would carry an extra error term that could break the sparsity assumption, so validating on noisy estimates is the most direct test."],"forward_implications":["At the same feedback overhead, the learned VQ codec on angle parameters improves reconstruction NMSE over the IEEE 802.11 standard quantizer, with the angle-difference strategy adding roughly 4 dB of NMSE gain at 576 bits on the DeepMIMO dataset.","The unified VQ method outperforms both the naive angle-difference feedback and the parallel VQ method while using only one decoder and no bit-allocation hyperparameter, reducing AP computation.","The CSI refinement module improves NMSE over no refinement, and the full combination (angle-difference feedback plus refinement) yields the highest net throughput in the simulated 802.11be system, about 64.7 Mb/s at 577 feedback bits versus about 19.4 Mb/s for standard T0 feedback at 9984 bits.","The reported gains persist across STA speeds from 0.4 to 2.8 m/s and across 20–320 MHz bandwidth, indicating the mechanism does not rely on massive-MIMO angular-delay sparsity.","On the Wi-MIR real-measurement dataset, the proposed framework also outperforms the standard feedback, supporting the claim beyond purely synthetic channels."],"supporting_citations":[{"why":"Defines the IEEE 802.11be system, the EVM-to-rate table used for throughput, and the feedback protocol that serves as the baseline and scenario.","marker":"[1]"},{"why":"Specifies the Givens-rotation angle feedback (φ, ψ) format that the paper uses as its encoder input and as the standard baseline.","marker":"[2]"},{"why":"Supplies the VQ-VAE training loss (distance, codebook, commitment) and the stop-gradient operator used to train encoder, codebook, and decoder jointly.","marker":"[15]"},{"why":"Reports that VQ-based DL CSI feedback outperforms scalar-quantization feedback in cellular MIMO, motivating the trainable VQ module in Wi-Fi.","marker":"[16]"},{"why":"Prior DL-based Wi-Fi CSI feedback method using hyperparameterized preprocessing and separate FCNs; used as a comparison baseline.","marker":"[27]"},{"why":"Prior CNN-based Wi-Fi CSI feedback that compresses the right singular matrix; used as a comparison baseline.","marker":"[28]"},{"why":"Wi-MIR real Wi-Fi channel dataset used for evaluation, providing real measured angle parameters with temporal correlation.","marker":"[37]"},{"why":"SimVP video prediction architecture adopted as the CSI refinement module to refine current angles from past frames.","marker":"[40]"},{"why":"DeepMIMO I3 synthetic dataset generated per IEEE 802.11be, used for training and testing.","marker":"[41]"}],"fun_headline_variants":["Neural CSI codec uses angle history, beats 802.11","Wi-Fi feedback learns from channel time correlation","DL shrinks CSI bits with angle-difference trick","Temporal-aware AI codec cuts Wi-Fi feedback cost","Vector-quantized CNN outdoes standard CSI feedback"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire pipeline assumes the STA's channel estimate is perfect; simulations feed noiseless channel matrices into SVD and angle extraction, so the learned codebooks and the sparse-difference assumption have not been tested against real channel estimation error.","fun_headline_variants_meta":{"raw":{"variants":["Neural CSI codec uses angle history, beats 802.11","Wi-Fi feedback learns from channel time correlation","DL shrinks CSI bits with angle-difference trick","Temporal-aware AI codec cuts Wi-Fi feedback cost","Vector-quantized CNN outdoes standard CSI feedback"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000311,"raw_usage":{"total_tokens":1820,"prompt_tokens":1043,"completion_tokens":777,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":659,"completion_tokens_details":{"reasoning_tokens":698}},"tokens_in":659,"tokens_out":777,"duration_ms":8435,"temperature":1.0,"reasoning_tokens":698,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T12:50:40.953222+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain and evaluate the same framework on the same DeepMIMO and Wi-MIR datasets but with channel estimates corrupted by the estimator used in NDP-based 802.11 acquisition, such as least-squares with the same pilot structure and receiver noise; if, at equal feedback overhead, the NMSE gap over the standard 802.11 quantizer disappears or the angle-difference mode rarely triggers, the claim of consistent gains is falsified.","supporting_citations":[{"cited_title":"Part 11: Wireless LAN medium access control (MAC) and physical Layer (PHY) specifications. Amendment 8: Enhancements for extremely high throughput,","cited_arxiv_id":null,"evidence_quote":"Defines the IEEE 802.11be system, the EVM-to-rate table used for throughput, and the feedback protocol that serves as the baseline and scenario."},{"cited_title":"Part 11: Wireless LAN medium access control (MAC) and physical layer (PHY) specifications. Amendment 4: Enhancements for very high throughput for operation in bands below 6GHz,","cited_arxiv_id":null,"evidence_quote":"Specifies the Givens-rotation angle feedback (φ, ψ) format that the paper uses as its encoder input and as the standard baseline."},{"cited_title":"Neural discrete representation learning,","cited_arxiv_id":null,"evidence_quote":"Supplies the VQ-VAE training loss (distance, codebook, commitment) and the stop-gradient operator used to train encoder, codebook, and decoder jointly."},{"cited_title":"Vector quantization for deep-learning- based CSI feedback in massive MIMO systems,","cited_arxiv_id":null,"evidence_quote":"Reports that VQ-based DL CSI feedback outperforms scalar-quantization feedback in cellular MIMO, motivating the trainable VQ module in Wi-Fi."},{"cited_title":"LB-SciFi: Online learning-based channel feedback for MU-MIMO in wireless LANs,","cited_arxiv_id":null,"evidence_quote":"Prior DL-based Wi-Fi CSI feedback method using hyperparameterized preprocessing and separate FCNs; used as a comparison baseline."},{"cited_title":"Deep learning-based CSI feedback in Wi-Fi systems,","cited_arxiv_id":null,"evidence_quote":"Prior CNN-based Wi-Fi CSI feedback that compresses the right singular matrix; used as a comparison baseline."},{"cited_title":"Wi-MIR: A CSI dataset for Wi-Fi based multi-person interaction recognition,","cited_arxiv_id":null,"evidence_quote":"Wi-MIR real Wi-Fi channel dataset used for evaluation, providing real measured angle parameters with temporal correlation."},{"cited_title":"SimVP: Simpler yet better video prediction,","cited_arxiv_id":null,"evidence_quote":"SimVP video prediction architecture adopted as the CSI refinement module to refine current angles from past frames."}],"review_version":1}