{"id":"4cece1db-617b-4456-a1c7-49a089ef5814","arxiv_id":"2506.04113","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"CSILocal reduces federated CSI feedback communication by exchanging smashed data at model splitting boundaries instead of full model parameters, with modest training time savings.","lead":"This paper proposes a federated learning scheme for CSI feedback compression in massive MIMO systems, splitting the autoencoder so that the encoder and decoder head stay on user devices while only the decoder tail resides at the base station. The approach exchanges only small boundary activations and gradients instead of full model parameters, aiming to cut communication overhead and, with pipeline parallelism, reduce training time.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"CSILocal's reported 19.66B exchanged parameters appears to count only one boundary tensor in one direction; counting all forward/backward boundary tensors reverses the communication-efficiency claim.","rationale":"The reader's weakest assumption already flags the omitted BTS-to-UE smashed data; this is the same load-bearing issue and I agree. The math is concrete: the reported 19.66B corresponds to a single one-way boundary tensor, while the scheme requires four such tensors per sample per iteration. The corrected total is about 78.6B at the 9,600-iteration operating point, or 163.8B at 20,000 iterations, which reverses the claimed ranking against FedAvg and FedProx. This is a correctness problem in the paper's principal quantitative claim, not a disagreement with community consensus. The pipeline-parallelism results in Table I are separate and may stand, but they do not support the communication-efficiency conclusion. Since the paper provides no code or detailed accounting to disambiguate the metric, the current version should be rejected; a revised version with full bidirectional boundary-tensor accounting and bit-level transmission cost could be reconsidered.","tokens_in":8835,"tokens_out":8790,"duration_ms":80616,"concrete_test":"Recompute Figure 6 with cumulative communicated scalars defined as the sum, over all iterations and UEs, of: (1) encoder smashed data uplink, (2) decoder-tail output downlink, (3) backward gradient w.r.t. decoder-tail output uplink, and (4) backward gradient w.r.t. encoder output downlink, using the stated batch size, c1=c2=256, and N=10. Then locate the testing-loss level 1.3e-3 on the corrected curves. If CSILocal's corrected value exceeds FedAvg's 47.06B or FedProx's 38.24B at that loss, Section IV.D's central comparison is invalid. An independent cross-check: derive the paper's 19.66B from the hyper-parameters; if it matches one direction times 9,600 iterations, the metric is under-specified.","verdict_should_be":"REJECT","load_bearing_attack":"The central claim is that CSILocal needs only 19.66B exchanged parameters versus 47.06B (FedAvg) and 38.24B (FedProx) for the same testing loss (Section IV.D). The paper's own Remark 2 states that, per sample, each iteration exchanges four boundary tensors: encoder output (dimension c1) uplink, decoder-tail output (c2) downlink, and two matching gradient tensors in the backward pass. With N=10, per-UE mini-batch 800, and c1=c2=256 (the stated default), one forward boundary tensor alone contributes 10*800*256*20,000 = 40.96B scalars. All four boundary tensors give 163.84B. The quoted 19.66B equals 10*800*256*9,600, i.e., exactly one scalar per sample per iteration over 9,600 iterations. Thus the headline advantage appears to be an artifact of counting a single one-way tensor against full-model parameter counts for the baselines. Even the downlink decoder-tail output alone at 20,000 iterations exceeds FedProx's 38.24B. No code or bit-level accounting is provided, so the discrepancy cannot be attributed to quantization or compression of boundary tensors.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CSILocal, a split-learning/federated training method for deep-learning-based CSI feedback in massive MIMO. The autoencoder is divided into an encoder and decoder head kept at each UE and a decoder tail kept at the BTS; UEs and the BTS exchange only boundary activation and gradient tensors rather than full model parameters. The decoder tail is further split across multiple BTS GPUs with pipeline parallelism to reduce wall-clock training time. Experiments on CsiNet-style data compare CSILocal with FedAvg, FedProx, and personalized variants under indoor, outdoor, and non-IID settings, reporting testing loss versus exchanged parameters and per-iteration pipeline timings.","tokens_in":9068,"tokens_out":5586,"duration_ms":53590,"significance":"If the communication-efficiency claim were correct, the paper would offer a practical way to train CSI feedback networks in federated settings with reduced uplink traffic, and the pipeline parallelism at the BTS is a sensible engineering idea. The paper does provide numerical comparisons with six benchmark algorithms, includes a non-IID data setup, and reports measured pipeline timings. However, the central quantitative claim is undermined by an inconsistent accounting of exchanged tensors: the reported 19.66 billion exchanged parameters does not match the paper's own description of the boundary tensors exchanged per iteration. The pipeline speedup percentages in Table I are also miscalculated. Because the main contribution is communication efficiency, these issues are load-bearing and the manuscript in its current form does not support its principal claims.","major_comments":[{"comment":"The headline number \"19.66 billion exchanged parameters\" is not consistent with the paper's own specification. Remark 2 states that for each sample and iteration, four boundary tensors are exchanged: the encoder output of dimension c1, the decoder-tail output of dimension c2, and two matching gradient tensors in the backward pass. With N=10 UEs, a per-UE mini-batch size of 800, c1=c2=256, and K=20,000 iterations (Section IV.A), a single forward boundary tensor alone carries 10 x 800 x 256 x 20,000 = 4.096 x 10^10 scalars, and all four boundary tensors carry 1.6384 x 10^11 scalars. The reported 19.66 x 10^9 equals 10 x 800 x 256 x 9,600, i.e., one scalar per sample per UE for only 9,600 iterations, which does not correspond to any stated counting rule. Even the downlink decoder-tail output alone over 20,000 iterations is 4.096 x 10^10 scalars, exceeding the 3.824 x 10^10 claimed for FedProx. No quantization, compression, or subsampling of boundary tensors is described, so the discrepancy cannot be attributed to such mechanisms. The comparison against 47.06 billion and 38.24 billion parameters is therefore unsupported.","section":"Section IV.D, Remark 2"},{"comment":"The communication-cost metric counts numbers of exchanged parameters or scalars but does not account for actual bits or wireless resource consumption. In a real link, every exchanged tensor, including the BTS-to-UE transmission of the decoder-tail output [F2 composed with F_{n,1}(H_{n,m})], consumes time-frequency resources and energy. The paper's metric assigns zero communication cost to this downlink transmission and does not include the overhead of digital transmission even though Section II.B assumes digital transmission under channel capacity. Since the paper's central contribution is \"communication efficiency,\" the comparison should report total uplink and downlink bits per round with a stated precision or quantization format, and should include the decoder-tail output and backward gradient tensors in the accounting.","section":"Section II.B and Section IV.D"},{"comment":"The claimed wall-clock reductions are computed with the wrong denominator. For indoor data, mini-batch 800, and encoding dimension 256, the duration drops from 0.1160 s to 0.0952 s per iteration; the relative reduction is (0.1160 - 0.0952)/0.1160 = 17.9%, not the stated 21.85%. The paper uses the pipeline time as the denominator, which inflates the percentages. The same issue affects the other entries, e.g., 0.0583 to 0.0560 is a 3.9% reduction, not 4.11%. The conclusion that pipeline parallelism reduces training time is directionally correct, but the reported magnitudes are inaccurate and should be corrected.","section":"Section IV.B, Table I"}],"minor_comments":[{"comment":"There are several typos and textual slips: \"calcualtes\" should be \"calculates,\" \"normlized\" should be \"normalized,\" and Section IV.D contains the duplicated phrase \"can can.\"","section":"Algorithm 1, line 8; Section IV.A"},{"comment":"The term \"smashed data\" is used repeatedly but is never formally defined; the authors should define it as the boundary activation tensors exchanged at the split points, and consistently use that terminology.","section":"Section II.B"},{"comment":"The local objective f_n(θ_{n,1}, θ_2, θ_{n,3}; H_{n,k}) uses index k internally while the outer sum runs over m; please align the indexing to avoid confusion.","section":"Equation (2)"},{"comment":"The relationship between mini-batches, micro-batches, and the four-GPU pipeline is described only briefly; a clearer statement of how data are divided and how gradient aggregation across micro-batches is performed would improve reproducibility.","section":"Section III and Figure 3"}],"recommendation":"reject","confidential_remarks":"The central claim of communication efficiency is contradicted by the manuscript's own Remark 2 and hyperparameter settings. Counting all four boundary tensors that the algorithm actually exchanges gives 163.84 billion scalars, which is larger than the FedAvg and FedProx baseline totals. This is not a minor presentation flaw but an invalidation of the main contribution. The pipeline timing percentages are also miscalculated. I would not encourage a revision unless the authors can provide a corrected bit-level accounting that preserves a meaningful communication advantage, which would require substantially new experiments and possibly a different method."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Yeah, I read CSILocal. The headline claim—19.66B exchanged parameters versus 47.06B for FedAvg—doesn't survive contact with the paper's own Remark 2. Remark 2 says each iteration exchanges four boundary tensors per sample: encoder output up, decoder-tail output down, and two gradient tensors in the backward pass. With N=10, batch 800, c1=c2=256, one forward tensor alone is 10*800*256*20,000 = 40.96B over the full 20k iterations. The quoted 19.66B works out to exactly one tensor per sample for about 9,600 iterations. So they counted one direction, one tensor, and compared that against full-model parameter counts for the baselines. Count all four tensors and the total is around 78B at 9,600 iterations, or 164B at 20k—worse than FedAvg's 47.06B. The central communication-efficiency claim is an artifact.\n\nThat's a shame because the paper has some genuine content. The model-splitting architecture for CSI feedback is a reasonable extension of split learning to a domain where it hasn't been applied. The pipeline parallelism at the BTS is real, and the wallclock measurements (4% to 22% speedup depending on batch size) are concrete. The convergence plots across indoor/outdoor/non-IID are useful. The authors also cite the relevant split-learning and personalized-FL literature.\n\nBut the accounting error is load-bearing. The paper's own Remark 2 contradicts the numbers in Section IV.D. Beyond that, the evaluation has other gaps: no error bars, a single operating point (testing loss 1.3e-3) for the big comparison, idealized assumptions about identical UEs and error-free digital transmission, and no code. The pipeline speedup is measured on a small scale (2 micro-batches).\n\nWho's this for? Researchers working on communication-efficient federated CSI feedback would want to look at the system design, but they'd have to redo the communication accounting before trusting any of the conclusions. I would not cite this as evidence of communication efficiency. I'd still send it to a serious referee, because the error is subtle enough that an expert could catch it and the authors might have a defensible version (e.g., with quantization or asymmetric counting) that could be fixed. But my own verdict: the main claim fails as written.","headline":"CSILocal's headline communication saving is an artifact of counting one boundary tensor against full-model baselines; the paper has a useful system design but its central claim reverses when you count what its own Remark 2 lists.","tokens_in":9603,"tokens_out":5544,"would_cite":false,"duration_ms":49764,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Splitting the CSI-feedback autoencoder between user devices and the base station lets federated training reach the same test loss with less than half the exchanged parameters.","keywords":["CSI feedback","federated learning","model splitting","communication efficiency","pipeline parallelism","massive MIMO","autoencoder compression"],"falsifier":"Measure the true bytes and wall-clock time on a live wireless testbed with heterogeneous user devices, counting both uplink smashed data and downlink decoder-tail outputs plus protocol overhead; if at equal test loss the total transmitted bytes and training latency are not below the full-parameter-exchange baselines, the communication-efficiency claim is falsified.","tokens_in":8642,"feed_emoji":"📡","tokens_out":6421,"duration_ms":55413,"temperature":0.7,"pith_summary":"This paper tries to prove that federated training of a CSI-feedback autoencoder can be made communication-efficient by splitting the model so each user device keeps the encoder and decoder head while the base station holds the decoder tail. Instead of exchanging full model parameters, devices send only the small smashed-data tensors at the split boundaries during forward and backward passes. The authors report that this reaches the same testing loss (about 1.3e-3) with 19.66 billion exchanged parameters, versus 47.06 billion for FedAvg and 38.24 billion for FedProx. They also show that splitting the decoder tail across multiple GPUs at the base station cuts per-iteration wall-clock time. If correct, this lowers the communication bottleneck that limits federated CSI feedback in massive MIMO systems.","feed_headline":"Model splitting halves federated CSI-feedback traffic","feed_subtitle":"UEs swap only smashed data with a shared base-station decoder, matching test loss with 58% fewer parameters.","key_machinery":"The load-bearing object is the three-way model split with boundary data exchange: encoder $F_{n,1}$ and decoder head $F_{n,3}$ stay at each UE, and decoder tail $F_2$ sits at the BTS. During forward passes the UEs upload smashed data $F_{n,1}(H_{n,m})$ and receive $F_2(F_{n,1}(H_{n,m}))$; during backward passes they exchange the corresponding gradients. Because the boundary dimensions satisfy $c_1 \\ll \\min\\{d_1,d_2\\}$ and $c_2 \\ll \\min\\{d_2,d_3\\}$, the communication volume is governed by the smashed-data sizes rather than the model sizes. The pipeline module further splits $F_2$ into sub-models on multiple GPUs to overlap forward and backward computation across micro-batches.","core_discovery":"The central claim is that CSILocal, a model-splitting algorithm for federated CSI feedback, significantly reduces the number of parameters exchanged between UEs and the BTS during training while keeping reconstruction accuracy essentially unchanged. The CSI autoencoder is divided into three modules: the encoder $F_{n,1}$ and decoder head $F_{n,3}$ reside at each UE, and the decoder tail $F_2$ is shared at the BTS. Communication is limited to the boundary activation tensors and their gradients, whose dimensions $c_1$ and $c_2$ are chosen much smaller than the parameter counts of the modules. In numerical experiments with 10 UEs, CSILocal reaches a testing loss of about $1.3\\times10^{-3}$ with 19.66 billion exchanged parameters, compared with 47.06 billion for FedAvg and 38.24 billion for FedProx. A pipeline module that splits the decoder tail across GPUs reduces the wall-clock duration per iteration by up to about 22% at batch size 800.","pith_inferences":["The communication accounting counts exchanged parameters, not transmitted bits; a fairer comparison would include the downlink smashed data sent from the BTS to each UE, digital-modulation overhead, and retransmissions, which could shrink or reverse the reported gap.","The privacy argument that raw CSI never leaves the UE does not automatically cover the smashed-data tensors, which are learned features of the channel and may encode location or identity information; a membership-inference test on the boundary activations would probe this.","The same split-and-exchange pattern could be applied to other split points or to hierarchical aggregators, and comparing against gradient-compression baselines (quantization, sparsification) would show whether model splitting or compression is the dominant source of the savings."],"forward_implications":["At equal testing loss, CSILocal exchanges about 58% fewer parameters than FedAvg and about 49% fewer than FedProx in the reported 10-UE setup.","The communication cost per iteration scales with the boundary dimensions $c_1$ and $c_2$, not with the total model size, so larger models do not automatically imply heavier federated traffic.","Pipeline parallelism in the decoder tail shortens per-iteration wall-clock time, with the reported reduction growing with mini-batch size and number of UEs.","Higher encoding dimensions compensate for non-IID-induced reconstruction degradation, at the price of larger smashed-data exchanges.","Keeping the original CSI matrices at each UE preserves local data privacy in the sense that raw CSI is not transmitted."],"supporting_citations":[{"why":"Supplies the CsiNet encoder architecture and dataset used for the autoencoder and experiments.","marker":"[10]"},{"why":"Defines the communication-efficient personalized federated edge learning baseline for CSI feedback that CSILocal is compared against.","marker":"[23]"},{"why":"Provides the gossip-training baseline that exchanges full parameters with neighbouring UEs, motivating the full-parameter bottleneck.","marker":"[24]"},{"why":"Supplies the GPipe pipeline-parallelism mechanism adapted to split the decoder tail across GPUs at the BTS.","marker":"[27]"},{"why":"Defines the partial-model-personalization benchmarks (FedAvgPer, FedProxPer, FedGradPer) used as baselines.","marker":"[29]"},{"why":"Provides the Adam optimizer and its convergence properties used for updating all model parts.","marker":"[26]"},{"why":"Justifies the digital-transmission assumption that the UE-BTS link has negligible error rate, underpinning the straggler assumption.","marker":"[25]"}],"fun_headline_variants":["Model splitting cuts CSI-FL parameter traffic by 58%","Federated CSI feedback: 58% fewer parameters via split","CSI-FL: split model, halve the communication load","Share the decoder, cut CSI-FL training chatter by 58%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central assumption is that all user devices have identical processing capabilities, so stragglers can be ignored, and that communication cost is faithfully measured by the number of exchanged parameters rather than actual transmitted bits, latency, or the energy of sending smashed data back to devices.","fun_headline_variants_meta":{"raw":{"variants":["Model splitting cuts CSI-FL parameter traffic by 58%","Federated CSI feedback: 58% fewer parameters via split","CSI-FL: split model, halve the communication load","Share the decoder, cut CSI-FL training chatter by 58%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000401,"raw_usage":{"total_tokens":2055,"prompt_tokens":869,"completion_tokens":1186,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":485,"completion_tokens_details":{"reasoning_tokens":1112}},"tokens_in":485,"tokens_out":1186,"duration_ms":11186,"temperature":1.0,"reasoning_tokens":1112,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T10:45:33.281897+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the true bytes and wall-clock time on a live wireless testbed with heterogeneous user devices, counting both uplink smashed data and downlink decoder-tail outputs plus protocol overhead; if at equal test loss the total transmitted bytes and training latency are not below the full-parameter-exchange baselines, the communication-efficiency claim is falsified.","supporting_citations":[{"cited_title":"Communication-efficient personalized federated edge learning for massive MIMO CSI feedback,","cited_arxiv_id":null,"evidence_quote":"Defines the communication-efficient personalized federated edge learning baseline for CSI feedback that CSILocal is compared against."},{"cited_title":"Gpipe: Efficient training of giant neural networks using pipeline parallelism,","cited_arxiv_id":null,"evidence_quote":"Supplies the GPipe pipeline-parallelism mechanism adapted to split the decoder tail across GPUs at the BTS."},{"cited_title":"Federated learning with partial model personalization,","cited_arxiv_id":null,"evidence_quote":"Defines the partial-model-personalization benchmarks (FedAvgPer, FedProxPer, FedGradPer) used as baselines."},{"cited_title":"A simple convergence proof of Adam and Adagrad,","cited_arxiv_id":null,"evidence_quote":"Provides the Adam optimizer and its convergence properties used for updating all model parts."},{"cited_title":"Wireless federated learning over resource-constrained networks: Digital versus analog transmissions,","cited_arxiv_id":null,"evidence_quote":"Justifies the digital-transmission assumption that the UE-BTS link has negligible error rate, underpinning the straggler assumption."}],"review_version":1}