Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

GNN-based Precoder Design and Fine-tuning for Cell-free Massive MIMO with Real-world CSI

T0 review · 3 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Fine-tuning a GNN precoder on real-world channel measurements, with the first four of eight layers frozen, yields an 8.2 bits/channel use sum-rate gain over the synthetic-only pretrained model at 20 dB SNR.

desk verdict A credible first step on sim-to-real GNN precoding with a public CSI dataset, but the headline gain is inflated by test-set model selection and tuple-level leakage in the four-user dataset. read the letter →

arxiv 2505.08788 v1 pith:PMKYEF6Y submitted 2025-05-13 eess.SP

Tianzheng Miao , Thomas Feys , Gilles Callebaut , Jarne Van Mulders , Emanuele Peschiera , Md Arifur Rahman , François Rottenberg This is my paper · ORCID
classification eess.SP
keywords cell-freemassiveMIMOgraphneuralnetworkprecodingtransferlearningfine-tuningreal-worldCSIlayerfreezingsumrate
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper claims that a graph neural network precoder trained on synthetic channel data can be made to work on real-world channels by fine-tuning on a limited set of measured channel vectors, with the first four of eight layers frozen during fine-tuning. In the four-user experiment on a 33-access-point testbed, fine-tuning improves sum rate by about 8.2 bits per channel use at 20 dB transmit SNR over the pretrained model, a relative gain of roughly 15.7 percent, and reduces the gap to zero-forcing precoding from 15.7 to 7.5 bits per channel use. The result matters because learned precoders are usually validated only in simulation, and this is a concrete demonstration that transfer learning can carry one from synthetic to measured channels.

What carries the argument

The load-bearing object is an eight-layer edge-centric graph neural network operating on a bipartite graph whose vertices are access points and users and whose edges carry the CSI as attributes. Each layer updates an edge representation by combining its previous embedding with mean-aggregated messages from the two endpoint nodes, applies a LeakyReLU nonlinearity, and the final output is power-normalized and trained unsupervised to maximize sum rate. The transfer mechanism is layer freezing: the first four layers keep their pretrained weights during fine-tuning and only the remaining layers are updated, which the authors find optimal after comparing all freezing depths from zero to eight.

What would settle it

Collect true simultaneous two-user and four-user CSI on the same 33-access-point testbed, run the same pretrain-and-Freeze-4 fine-tuning protocol, and compare sum rates; if the fine-tuned GNN no longer gains roughly 8.2 bits per channel use over the pretrained model at 20 dB SNR, or no longer cuts the zero-forcing gap from 15.7 to about 7.5 bits per channel use, the reported gain is an artifact of the constructed dataset.

Watch

Extended reading notes

Core claim

The central claim is that unsupervised pretraining on synthetic CSI followed by unsupervised fine-tuning on real CSI makes an edge-centric GNN precoder generalize to a physical cell-free massive MIMO environment, whereas the same model trained only on synthetic CSI degrades when evaluated on real measurements. In the four-user setting, the fine-tuned model outperforms the pretrained model by 8.2 bits per channel use at 20 dB SNR and narrows the gap to zero-forcing from 15.7 to 7.5 bits per channel use; a model trained from scratch on the real data alone does not match this performance. The authors attribute the improvement to pretraining providing a favorable initialization and to freezing early layers preserving generalizable representations while later layers specialize to the real channel statistics.

Load-bearing premise

The four-user real-world data are not actual simultaneous four-user measurements; they are built by pairing single-user channel vectors and keeping only the 44 strongest channels, so the reported gain assumes those constructed matrices behave like genuine multi-user propagation.

Editorial extensions

If this is right

  • A synthetic-to-real transfer recipe exists for GNN precoding: pretrain on simulated channels, then fine-tune on a modest set of measured channels.
  • Freezing early layers outperforms full fine-tuning in this setup, so preserving pretrained representations matters more than full adaptation when real data are scarce.
  • The fine-tuned GNN remains 7.5 bits per channel use below zero-forcing on real four-user channels, so learned precoding is closer to, but not yet a replacement for, zero-forcing.
  • Fine-tuning improves real-data performance at the cost of some synthetic-data performance, confirming a domain-adaptation tradeoff rather than universal improvement.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The authors leave implicit that the optimal freezing depth likely depends on the size of the real dataset and the severity of the distribution shift, so the Freeze-4 choice should be re-tuned rather than reused in other deployments.
  • Because the four-user real dataset is constructed by pairing single-user channel vectors and keeping only the 44 strongest channels, the reported gain may differ on genuinely simultaneous multi-user measurements; that is a testable extension, not a claim in the paper.
  • The same pretrain-then-freeze pipeline could plausibly transfer to other learned MIMO components, such as channel estimation or RIS beamforming, which face the same sim-to-real gap.
  • A natural next experiment is to vary the number of real channel samples used for fine-tuning; the paper uses one fixed real dataset and does not report how the 8.2 bits per channel use gain scales with data volume.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes a graph neural network (GNN) precoder for cell-free massive MIMO, first pre-trained on synthetic channels generated with an Indoor Hotspot path-loss model and Rayleigh fading, then fine-tuned on real CSI collected from the Techtile testbed. To adapt the pretrained model, the authors freeze the first l of eight GNN layers and retrain the remaining layers on real CSI, evaluating every l in {0,...,8}. They construct two-user and four-user datasets by forming tuples of single-user measured channel vectors. The reported results claim that the Freeze-4 model gains approximately 8.2 bit/channel use over the pretrained model at 20 dB SNR on real CSI (about 15.7% relative improvement) and reduces the gap to zero-forcing from 15.7 to 7.5 bit/channel use. The paper also compares against zero-forcing, conjugate beamforming, and a randomly initialized model trained directly on real data.

Significance. If the findings were obtained under a leakage-free evaluation, the paper would provide a useful empirical data point on sim-to-real transfer for GNN-based precoding, with a publicly stated dataset and an unsupervised objective matched to sum-rate maximization. The central direction and the choice of baselines (ZF, CB, no-pretraining) are sensible, and the layer-freezing idea is clearly motivated by limited real data. However, the current evaluation protocol has a load-bearing flaw: the freeze depth is selected on the same test set used for the headline numbers, and the constructed real-world datasets reuse the same measured single-user channel vectors in training and testing. These issues mean the specific quantitative claims are not yet established, even though the general direction may be correct.

major comments (3)
  1. [§IV.A and §V.C (Fig. 6(b))] The four-user real-world dataset is assembled from 4-tuples of the same 44 strongest single-user channel vectors, and the random 80/10/10 split is performed at the tuple level. Because each of the 44 base vectors appears in many training and test tuples, the same measured channel vectors occur on both sides of the split. The reported 8.2 bit/channel use gain at 20 dB SNR and the reduction of the ZF gap from 15.7 to 7.5 bits therefore reflect interpolation over a very small set of measured channels rather than generalization to unseen user positions. The evaluation should be repeated with a position-level split in which the set of single-user positions used for fine-tuning is disjoint from the set used for testing.
  2. [§V.A] The freeze-depth parameter l is selected by exhaustively evaluating all l in {0,...,8} on the same four-user real-world CSI test set that is later used to report every headline performance number. This makes the choice of Freeze-4 and the 8.2 bit/channel use gain a selected maximum over nine configurations evaluated on the test set, which biases the result upward. The freezing strategy should be chosen on the validation split or via nested cross-validation, with the final test set used only once.
  3. [§IV.A and §V.C] The constructed multi-user datasets concatenate single-user channel vectors measured at different positions and times, and the four-user set keeps only the 44 largest-norm single-user channels. Such matrices do not contain realistic inter-user correlation, joint shadowing/fading, or user scheduling effects, and the strong-channel selection restricts the test distribution to high-gain users. The statement in Section V.C that fine-tuning learns from real-world interference patterns is therefore not supported by this dataset construction; the results should be framed as a proof-of-concept on constructed multi-user samples, not as validation on genuine multi-user real-world measurements.
minor comments (5)
  1. [Eq. (10)] Equation (10) is typeset as a single equation with two branches but no clear delimiter; it should be split into two separate labeled equations for CB and ZF.
  2. [References] References [7] and [8] are identical; the duplicate should be removed and the reference list should be checked for other duplicates.
  3. [§IV.A] The dataset availability statement mentions 'Real-world CSI Dataset' without an actual URL or repository identifier; a working link should be provided.
  4. [§V.C] The phrase 'corresponding to a relative improvement of approximately 15.7%' is ambiguous because the immediately preceding numbers are the ZF gap reduction (15.7 to 7.5 bits), while the 15.7% refers to the sum-rate gain of the fine-tuned model over the pretrained model; please clarify this wording.
  5. [§V.A and Fig. 4] If Freeze 8 freezes all layers, the model is identical to the pretrained network, so the relation between the Freeze 8 curve in Fig. 4 and the 'GNN Pretrained (real CSI)' curve in Fig. 5 should be stated explicitly to avoid apparent inconsistency.

Circularity Check

2 steps flagged · score 4.0 of 10

The reported 8.2 bit/channel-use fine-tuning gain is evaluated on the same real-CSI four-user test set used to select the freeze depth, and that test set is constructed from the same 44 measured single-user channel vectors that appear in the training tuples, so the headline is a leaky, post-selection estimate rather than an independent sim-to-real prediction.

  1. fitted input called prediction [Section V.A (Freezing Strategy and Fine-Tuning Performance), Fig. 4; reported in Section V.C, Fig. 6(b)]
    "To determine the most effective freezing strategy for subsequent fine-tuning, we evaluated various configurations using real-world CSI data from a four-user scenario as the test set, with the learning rate kept consistent with that used during pretraining. As illustrated in Fig. 4, freezing the first 4 layers achieves the best performance. This configuration strikes a balance between retaining the pretrained model's prior knowledge and maintaining sufficient adaptability to the new real-world dataset."

    The freeze depth l is not chosen on a held-out validation set; it is selected by exhaustive evaluation of all l in {0,...,8} on the same real-world four-user test set that later produces the reported 8.2 bits/channel-use improvement in Fig. 6(b). The selection criterion (maximizing sum rate on the four-user real test set) and the reported outcome (the sum-rate gain of the selected Freeze-4 model on that same test set) are the same function of the same data. The headline number is therefore a post-selection, test-set-optimistic estimate rather than an independent prediction of fine-tuning benefit, because Freeze-4 was chosen specifically because it performed best on exactly the examples used to measure the gain.

  2. other [Section IV.A, Data Collection and Dataset Preparation]
    "we select the top 44 single-user samples based on channel strength ∥hi∥. This selection ensures a sufficiently large number of unique four-user combinations, as 44 4 = 135751≥124750 . From these, we randomly sample 124 750 combinations without replacement to match the size of the two-user dataset. ... For model training and evaluation, both the two-user and four-user datasets are partitioned into training, validation, and testing subsets, containing 80%, 10%, and 10% of the total samples."

    All four-user test samples are unordered 4-tuples drawn from the same pool of 44 measured single-user channel vectors used to form training tuples, and the train/test split is random at the tuple level. Since 124,750 of the 135,751 possible 4-tuples are used, nearly every one of the 44 base vectors appears in both training and test tuples. Consequently the 'unseen' test data are not unseen channels: each test input is a recombination of channel vectors the model has seen during fine-tuning. The headline 8.2 bit/channel-use gain can therefore be driven by memorization or interpolation over the same 44 measured vectors, rather than by generalization to new user positions.

full rationale

This is an empirical evaluation paper rather than a derivation, so there is no chain of equations in which a predicted quantity equals an input by algebra. The self-citations in the paper (notably [4] for the GNN architecture and [10] for the Techtile testbed) are supporting tools and are not load-bearing uniqueness theorems; citing one's own architecture and testbed is normal and does not by itself make the transfer-learning result circular. The central result, that fine-tuning a GNN on real CSI improves sum rate relative to the pretrained model, is a measurement, and measurements can in principle be valid even when prior architecture work is self-cited. However, two evaluation choices undermine the independence of the headline 8.2 bit/channel-use number. First, the layer-freezing hyperparameter was selected by evaluating all nine freeze depths on the same four-user real-CSI test set that is later used to report the improvement, so the reported gain is a post-selection estimate biased by test-set peeking. Second, the four-user real dataset is generated from only 44 measured single-user channel vectors, with a random split at the 4-tuple level, so training and test tuples share the same base channel vectors; the model can memorize or interpolate over those 44 vectors rather than generalizing to new positions. These issues make the claimed sim-to-real gain a leaky, partially circular estimate, but they do not make the result true by definition, since the fine-tuned model could still fail on the selected tuples. The appropriate score is therefore 4 rather than 6 or higher.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central claim rests on standard propagation assumptions, the validity of constructing multi-user CSI from single-user measurements, and two hand-chosen experimental parameters (freeze depth and top-44 selection). No new physical entities are introduced.

free parameters (2)
  • freeze-layer count l = 4 (chosen by exhaustive evaluation on the real-CSI test set)
    Selected in Section V.A by evaluating all l in {0,...,8} on the four-user real CSI test set, then used for all final results. This is a hyperparameter tuned on the evaluation data.
  • number of strongest single-user samples for four-user dataset = 44
    Chosen in Section IV.A so that C(44,4)=135751 provides at least 124750 four-user combinations. This selection biases the dataset toward strong channels.
assumptions (4)
  • domain assumption 3GPP InH NLOS path loss model (TR 38.901) is used for synthetic data generation.
    Invoked in Section II, Eq. (2). The synthetic pretraining relies on this model as a stand-in for real propagation.
  • domain assumption Small-scale fading is i.i.d. Rayleigh distributed for synthetic data.
    Stated in Section II: hm,k ~ CN(0,1). This is a standard but idealized assumption that the real data may violate.
  • domain assumption Constructed multi-user CSI from independent single-user measurements is representative of real multi-user channels.
    The real dataset in Section IV.A is built by pairing channel vectors from a single UE at different positions. This ignores inter-user correlation and simultaneous multi-user interference patterns present in actual deployments.
  • domain assumption Centralized precoding with perfect global CSI at the CPU and ideal fronthaul.
    The system model in Section II assumes the CPU has full CSI. Real fronthaul and estimation errors are not modeled.

how reviews work

0 comments
Cite this review

Pith. "Pith review of GNN-based Precoder Design and Fine-tuning for Cell-free Massive MIMO with Real-world CSI." pith.science (2026). https://pith.science/paper/PMKYEF6Y

@misc{pith2026250508788,
  author       = {Pith},
  title        = {Pith review of: GNN-based Precoder Design and Fine-tuning for Cell-free Massive MIMO with Real-world CSI},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PMKYEF6Y}},
  note         = {Machine review of arXiv:2505.08788}
}
read the original abstract

Cell-free massive MIMO (CF-mMIMO) has emerged as a promising paradigm for delivering uniformly high-quality coverage in future wireless networks. To address the inherent challenges of precoding in such distributed systems, recent studies have explored the use of graph neural network (GNN)-based methods, using their powerful representation capabilities. However, these approaches have predominantly been trained and validated on synthetic datasets, leaving their generalizability to real-world propagation environments largely unverified. In this work, we initially pre-train the GNN using simulated channel state information (CSI) data, which incorporates standard propagation models and small-scale Rayleigh fading. Subsequently, we finetune the model on real-world CSI measurements collected from a physical testbed equipped with distributed access points (APs). To balance the retention of pre-trained features with adaptation to real-world conditions, we adopt a layer-freezing strategy during fine-tuning, wherein several GNN layers are frozen and only the later layers remain trainable. Numerical results demonstrate that the fine-tuned GNN significantly outperforms the pre-trained model, achieving an approximate 8.2 bits per channel use gain at 20 dB signal-to-noise ratio (SNR), corresponding to a 15.7 % improvement. These findings highlight the critical role of transfer learning and underscore the potential of GNN-based precoding techniques to effectively generalize from synthetic to real-world wireless environments.

Figures

Figures reproduced from arXiv: 2505.08788 by the authors.

Figure 1
Figure 1. System model of the considered CF-mMIMO network. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Illustration of the graph representing the cell-free system [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Testbed setup (Left: Techtile environment with 33 APs on the ceiling and UE at floor. Right: Illustration of the hardware setup for each AP, deployed on the backside of the ceiling planks.) In the following, we describe the design and fine-tuning strategy of the proposed GNN-based precoder, shown in [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Comparison of different freezing strategies. Freeze [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: Sum-rate performance comparison for different precod [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: Sum-rate performance of the CF-mMIMO system with [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Learning to Quantize and Precode in Massive MIMO Systems for Energy Reduction: a Graph Neural Network Approach

    eess.SY 2025-07 conditional novelty 5.0 of 10

    A self-supervised GNN with straight-through Gumbel-softmax training performs non-linear quantized precoding, matching 3-bit MRT rate with 1-bit DACs in single-user massive MIMO, though the GNN processing power limits ...

Reference graph

Works this paper leans on

12 extracted references · 10 canonical work pages · cited by 1 Pith paper

  1. [1]

    Cell-free massive MIMO versus small cells,

    H. Q. Ngo, A. Ashikhmin, H. Yang, E. G. Larsson, and T. L. Marzetta, “Cell-free massive MIMO versus small cells,”IEEE Transactions on Wireless Communications, vol. 16, no. 3, pp. 1834–1850, 2017

    2017

  2. [2]

    Graph neural networks meet wireless communications: Motivation, applications, and future directions,

    M. Lee, G. Yu, H. Dai, and G. Y . Li, “Graph neural networks meet wireless communications: Motivation, applications, and future directions,” IEEE Wireless Communications, vol. 29, no. 5, pp. 12–19, 2022

    work page 2022

  3. [3]

    Learning to Reflect and to Beamform for Intelligent Reflecting Surface With Implicit Channel Estimation,

    T. Jiang, H. V . Cheng, and W. Yu, “Learning to Reflect and to Beamform for Intelligent Reflecting Surface With Implicit Channel Estimation,” IEEE Journal on Selected Areas in Communications, vol. 39, no. 7, pp. 1931–1945, Jul. 2021

    work page 1931

  4. [4]

    Toward energy-efficient massive MIMO: Graph neural network precoding for mitigating non- linear PA distortion,

    T. Feys, L. Van der Perre, and F. Rottenberg, “Toward energy-efficient massive MIMO: Graph neural network precoding for mitigating non- linear PA distortion,”IEEE Transactions on Cognitive Communications and Networking, vol. 11, no. 1, pp. 184–201, 2025

    work page 2025

  5. [5]

    DeepMIMO: A Generic Deep Learning Dataset for Millimeter Wave and Massive MIMO Applications,

    A. Alkhateeb, “DeepMIMO: A Generic Deep Learning Dataset for Millimeter Wave and Massive MIMO Applications,” Feb. 2019

    work page 2019

  6. [6]

    WAIR-D: Wireless AI research dataset,

    Y . Huangfu, J. Wang, S. Dai, R. Li, J. Wang, C. Huang, and Z. Zhang, “WAIR-D: Wireless AI research dataset,”arXiv preprint arXiv:2212.02159, 2022

    arXiv 2022

  7. [7]

    Transfer Learning Promotes 6G Wireless Communications: Recent Advances and Future Challenges,

    M. Wang, Y . Lin, Q. Tian, and G. Si, “Transfer Learning Promotes 6G Wireless Communications: Recent Advances and Future Challenges,” IEEE Transactions on Reliability, vol. 70, no. 2, pp. 790–807, Jun. 2021

    work page 2021

  8. [8]

    Transfer Learning Promotes 6G Wireless Communications: Recent Advances and Future Challenges,

    ——, “Transfer Learning Promotes 6G Wireless Communications: Recent Advances and Future Challenges,”IEEE Transactions on Reliability, vol. 70, no. 2, pp. 790–807, Jun. 2021

    work page 2021

Show all 12 references
  1. [9]

    Study on channel model for frequencies from 0.5 to 100 ghz (release 17),

    3GPP, “Study on channel model for frequencies from 0.5 to 100 ghz (release 17),” 3GPP, Tech. Rep. TR 38.901 V17.0.0, 2022, (March 2022)

    2022

  2. [10]

    Techtile: Open 6G R&D testbed for communica- tion, positioning, sensing, WPT and federated learning,

    G. Callebaut, J. Van Mulders, G. Ottoy, D. Delabie, B. Cox, N. Stevens, and L. Van der Perre, “Techtile: Open 6G R&D testbed for communica- tion, positioning, sensing, WPT and federated learning,” inProc. Joint European Conference on Networks and Communications & 6G Summit (Eu...

    2022

  3. [11]

    Quinonero-Candela, M

    J. Quinonero-Candela, M. Sugiyama, A. Schwaighofer, and N. D. Lawrence,Dataset Shift in Machine Learning. MIT Press, 2008

    2008

  4. [12]

    Optimal Design of Energy-Efficient Multi-User MIMO Systems: Is Massive MIMO the Answer?

    E. Bj ¨ornson, L. Sanguinetti, J. Hoydis, and M. Debbah, “Optimal Design of Energy-Efficient Multi-User MIMO Systems: Is Massive MIMO the Answer?”IEEE Transactions on Wireless Communications, vol. 14, no. 6, pp. 3059–3075, Jun. 2015

    2015

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.