pith. sign in

arxiv: 2511.11251 · v2 · submitted 2025-11-14 · 📡 eess.SP

Testbed Evaluation of AI-based Precoding in Distributed MIMO Systems

Tianzheng Miao , Thomas Feys , Gilles Callebaut , Jarne Van Mulders , Md Arifur Rahman , Fran\c{c}ois Rottenberg This is my paper

Pith reviewed 2026-05-17 22:28 UTC · model grok-4.3

classification 📡 eess.SP
keywords distributed MIMOAI precodinggraph neural networktestbed evaluationfine-tuningchannel state informationreciprocity calibration6G
0
0 comments X

The pith

Fine-tuning a graph neural network on real distributed MIMO channel data yields a 15.7 percent multi-user performance gain and near-maximum ratio transmission results on unseen positions.

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

The paper establishes a practical validation framework for AI-driven precoding in distributed MIMO systems by taking a pre-trained graph neural network model and adapting it with channel measurements gathered from a hardware testbed that includes reciprocity calibration. This adaptation produces measurable throughput improvements in both single-user and multi-user settings, closing much of the gap between simulation-trained models and real-world operation even when users appear at locations absent from the fine-tuning data. A sympathetic reader would care because most prior work on AI precoders remained confined to idealized mathematical models, leaving open whether such methods can handle hardware non-idealities and generalize outside the training set in actual 6G-style deployments.

Core claim

The central claim is that a pre-trained GNN-based precoder, after fine-tuning on real-world CSI collected from the Techtile distributed MIMO platform with hardware reciprocity calibration, delivers a 15.7 percent performance gain over the pre-trained version in the multi-user case and achieves near-MRT performance in the single-user case with less than 0.7 bits per channel use degradation from a total throughput of 5.19 bits per channel use, even at previously unseen user positions; additional measurements confirm that performance improves consistently with more real training samples and that end-to-end transmission shows coherent power focusing comparable to conventional MRT.

What carries the argument

A graph neural network that maps channel state information to precoding vectors, initially trained on simulated channels and then fine-tuned on real testbed CSI after hardware reciprocity calibration.

If this is right

  • Fine-tuning with real CSI improves performance under both interpolation and extrapolation to new user locations.
  • The approach exhibits data efficiency, producing further gains as the number of real-world training samples increases.
  • End-to-end over-the-air validation confirms coherent power focusing that matches conventional MRT behavior.
  • Hardware reciprocity calibration suffices to transfer the learned precoder from simulation to the physical testbed.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The results imply that AI precoders may reduce reliance on perfectly accurate analytical channel models in future cell-free or distributed 6G networks.
  • The same fine-tuning pipeline could be tested on larger arrays to check whether graph neural network scalability limits appear in practice.
  • The framework offers a template for validating other learned signal-processing blocks such as channel estimation or detection under real hardware conditions.

Load-bearing premise

Fine-tuning on CSI from the testbed produces a model that generalizes reliably to unseen positions and that the hardware reciprocity calibration fully captures real-world non-idealities without introducing unaccounted biases.

What would settle it

Collect throughput measurements at a fresh set of user positions after fine-tuning and compare against MRT; if the observed degradation substantially exceeds 0.7 bits per channel use from the 5.19 bits per channel use baseline, the generalization result would be falsified.

Figures

Figures reproduced from arXiv: 2511.11251 by Fran\c{c}ois Rottenberg, Gilles Callebaut, Jarne Van Mulders, Md Arifur Rahman, Thomas Feys, Tianzheng Miao.

Figure 1
Figure 1. Figure 1: System model of the considered D-MIMO network. Dis [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Left: The Techtile support structure hosting 140 tiles – Right: The back of three of such tiles, equipped with the default setup, i.e., [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: TDD-based deployment. The environment can be considered [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Sum-rate performance of the D-MIMO system with different [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Heatmap of extrapolation and interpolation performance gap [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Sample efficiency (shaded regions show variability across [PITH_FULL_IMAGE:figures/full_fig_p005_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Received power heatmaps in the target area under different precoding schemes and varying numbers of cooperating APs, each [PITH_FULL_IMAGE:figures/full_fig_p006_7.png] view at source ↗
read the original abstract

Distributed MIMO (D-MIMO) has emerged as a key architecture for future sixth-generation (6G) networks, enabling cooperative transmission across spatially distributed access points (APs). However, most existing studies rely on idealized channel models and lack hardware validation, leaving a gap between algorithmic design and practical deployment. Meanwhile, recent advances in artificial intelligence (AI)-driven precoding have shown strong potential for learning nonlinear channel-to-precoder mappings, but their real-world deployment remains limited due to challenges in data collection and model generalization. This work presents a framework for implementing and validating an AI-based precoder on a D-MIMO testbed with hardware reciprocity calibration. A pre-trained graph neural network (GNN)-based model is fine-tuned using real-world channel state information (CSI) collected from the Techtile platform and evaluated under both interpolation and extrapolation scenarios before end-to-end validation. Experimental results demonstrate a 15.7% performance gain over the pre-trained model in the multi-user case after fine-tuning, while in the single-user scenario the model achieves near-maximum ratio transmission (MRT) performance with less than 0.7 bits/channel use degradation out of a total throughput of 5.19 bits/channel use on unseen positions. Further analysis confirms the data efficiency of real-world measurements, showing consistent gains with increasing training samples, and end-to-end validation verifies coherent power focusing comparable to MRT.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 3 minor

Summary. The paper presents a framework for testbed evaluation of an AI-based precoder in distributed MIMO (D-MIMO) systems. It uses a pre-trained graph neural network (GNN) model that is fine-tuned on real-world channel state information (CSI) collected from the Techtile platform, incorporating hardware reciprocity calibration. Performance is assessed in both interpolation and extrapolation scenarios, followed by end-to-end hardware validation. Key reported outcomes include a 15.7% performance gain over the pre-trained model in the multi-user case after fine-tuning, and near-maximum ratio transmission (MRT) performance in the single-user case with less than 0.7 bits/channel use degradation out of 5.19 bits/channel use on unseen positions, along with analysis of data efficiency.

Significance. If the central empirical claims hold, this work is significant for providing hardware-validated evidence of AI-driven precoding in a real D-MIMO testbed, addressing the common gap between idealized simulations and practical 6G deployments. The demonstration of fine-tuning benefits using measured CSI, data efficiency trends, and coherent power focusing comparable to MRT offers concrete insights into model generalization and the value of hardware-in-the-loop approaches. The end-to-end validation on the Techtile platform strengthens the practical relevance.

major comments (3)
  1. [Abstract] Abstract and results presentation: The quantitative claims of a 15.7% performance gain in the multi-user case and less than 0.7 bits/channel use degradation (out of 5.19) in the single-user extrapolation scenario are stated without error bars, standard deviations across trials, or any statistical significance tests. This directly affects the reliability assessment of the generalization claims to unseen positions.
  2. [Experimental Setup / Evaluation] Data collection and evaluation protocol: Insufficient detail is provided on the spatial distribution and selection criteria for CSI collection points used in training versus interpolation/extrapolation test sets. Without evidence that training and test positions span independent spatial diversity (e.g., via explicit description of multipath correlation or minimum separation distances), the reported extrapolation performance risks being partly due to channel similarity rather than robust generalization.
  3. [Hardware Implementation] Hardware reciprocity calibration: The assumption that the calibration procedure fully captures non-idealities without introducing unaccounted biases is load-bearing for the throughput numbers, yet the manuscript provides limited quantitative validation or sensitivity analysis of residual calibration errors and their impact on the observed gains versus MRT.
minor comments (3)
  1. [Results] Clarify notation for throughput units and ensure consistent use of 'bits/channel use' throughout figures and text.
  2. [Introduction] Add more references to prior GNN-based precoding works and recent D-MIMO testbed studies to better contextualize the contribution.
  3. [Figures] Improve figure captions to explicitly state the number of independent measurements or runs underlying each plotted point.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed comments on our manuscript. These suggestions will help strengthen the presentation of our results and the description of our experimental methodology. We address each major comment point by point below, with plans to incorporate revisions in the updated version of the paper.

read point-by-point responses

  1. Referee: [Abstract] Abstract and results presentation: The quantitative claims of a 15.7% performance gain in the multi-user case and less than 0.7 bits/channel use degradation (out of 5.19) in the single-user extrapolation scenario are stated without error bars, standard deviations across trials, or any statistical significance tests. This directly affects the reliability assessment of the generalization claims to unseen positions.

    Authors: We agree that including variability metrics would improve the reliability assessment of the reported gains. The original manuscript focused on the primary measured outcomes from the Techtile testbed without explicitly reporting error bars or standard deviations across repeated trials. In the revised manuscript, we will add error bars derived from multiple independent CSI collection runs, report the associated standard deviations, and briefly discuss the consistency of the 15.7% gain and the sub-0.7 bits/channel use degradation. This addition will better support the generalization claims without altering the core findings. revision: yes

  2. Referee: [Experimental Setup / Evaluation] Data collection and evaluation protocol: Insufficient detail is provided on the spatial distribution and selection criteria for CSI collection points used in training versus interpolation/extrapolation test sets. Without evidence that training and test positions span independent spatial diversity (e.g., via explicit description of multipath correlation or minimum separation distances), the reported extrapolation performance risks being partly due to channel similarity rather than robust generalization.

    Authors: We thank the referee for highlighting the need for greater clarity on spatial aspects. The manuscript references the Techtile platform's CSI collection but does not provide exhaustive details on position selection or correlation analysis. In the revision, we will expand the experimental setup section to include a description of the measurement grid, explicit minimum separation distances between training and test positions (on the order of several wavelengths to promote spatial decorrelation), and supporting discussion of multipath characteristics observed in the environment. These additions will demonstrate that the extrapolation results reflect robust generalization rather than unintended channel similarity. revision: yes

  3. Referee: [Hardware Implementation] Hardware reciprocity calibration: The assumption that the calibration procedure fully captures non-idealities without introducing unaccounted biases is load-bearing for the throughput numbers, yet the manuscript provides limited quantitative validation or sensitivity analysis of residual calibration errors and their impact on the observed gains versus MRT.

    Authors: We acknowledge that additional quantitative validation of the calibration would address potential concerns about residual biases. The current manuscript describes the hardware reciprocity calibration procedure applied prior to fine-tuning and evaluation. To strengthen this section, the revised manuscript will incorporate quantitative results on post-calibration residual phase and amplitude errors, along with a sensitivity analysis showing how small calibration inaccuracies would affect the throughput gap relative to MRT. This will confirm that the observed performance gains and near-MRT behavior are attributable to the fine-tuned model rather than calibration artifacts. revision: yes

Circularity Check

0 steps flagged

No circularity in empirical testbed evaluation

full rationale

The paper reports experimental results from hardware measurements on the Techtile platform, including fine-tuning a pre-trained GNN on collected CSI and evaluating throughput gains on unseen positions against MRT baselines. No mathematical derivations, first-principles predictions, or equations are presented whose outputs reduce to the inputs by construction. All performance claims (e.g., 15.7% gain, <0.7 bits/channel use degradation) derive directly from end-to-end hardware tests rather than any self-referential fitting or self-citation chain. The work is self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on standard wireless assumptions plus the validity of the testbed measurements and model generalization; no free parameters or invented entities are explicitly introduced in the abstract.

axioms (1)
  • domain assumption Hardware reciprocity calibration accurately compensates for non-reciprocal effects in the distributed MIMO testbed.
    Invoked when describing the framework for implementing the AI-based precoder on the testbed.

pith-pipeline@v0.9.0 · 5572 in / 1399 out tokens · 41435 ms · 2026-05-17T22:28:42.759816+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

18 extracted references · 18 canonical work pages

  1. [1]

    Cell-Free Massive MIMO Versus Small Cells,

    H. Q. Ngo et al., “Cell-Free Massive MIMO Versus Small Cells,” IEEE Trans. Wireless Commun., Mar. 2017

    work page 2017

  2. [2]

    User-Centric Cell-Free Massive MIMO Net- works: A Survey of Opportunities, Challenges and Solutions,

    H. A. Ammar et al., “User-Centric Cell-Free Massive MIMO Net- works: A Survey of Opportunities, Challenges and Solutions,”IEEE Commun. Surveys Tuts., 2022

    work page 2022

  3. [3]

    Over-the-Air DPD and Reciprocity Calibration in Massive MIMO and Beyond,

    A. Sheikhi et al., “Over-the-Air DPD and Reciprocity Calibration in Massive MIMO and Beyond,”IEEE Wireless Commun. Lett., 2025

    work page 2025

  4. [4]

    Experimental Investigation of TDD Reciprocity-Based Zero-Forcing Transmit Precoding,

    P. Zetterberg, “Experimental Investigation of TDD Reciprocity-Based Zero-Forcing Transmit Precoding,”EURASIP, Dec. 2011

    work page 2011

  5. [5]

    On the Road to 6G: Visions, Requirements, Key Technologies, and Testbeds,

    C.-X. Wang et al., “On the Road to 6G: Visions, Requirements, Key Technologies, and Testbeds,”IEEE Commun. Surveys Tuts., 2023

    work page 2023

  6. [6]

    Distributed Neural Precoding for Hybrid mmWave MIMO Communications With Limited Feedback,

    K. Wei et al., “Distributed Neural Precoding for Hybrid mmWave MIMO Communications With Limited Feedback,”IEEE Commun. Lett., Jul. 2022

    work page 2022

  7. [7]

    StudyonArtificialIntelligence(AI)/MachineLearning(ML) for NR air interface (Release 19),

    3GPP,“StudyonArtificialIntelligence(AI)/MachineLearning(ML) for NR air interface (Release 19),” 3GPP, Tech. Rep. TR 38.843 V19.0.0, 2025, (September 2025)

    work page 2025

  8. [8]

    Callebaut et al.,Usrp b210 based synchronization measure- ments, GitHub repository, techtile-by-dramco/NI-B210-Sync, Ac- cessed: 2025-09-08

    G. Callebaut et al.,Usrp b210 based synchronization measure- ments, GitHub repository, techtile-by-dramco/NI-B210-Sync, Ac- cessed: 2025-09-08

    work page 2025

  9. [9]

    Multilayer feedforward networks are universal approximators,

    K. Hornik et al., “Multilayer feedforward networks are universal approximators,”Neural Networks, Jan. 1989

    work page 1989

  10. [10]

    Learning to Quantize and Precode in Massive MIMO Systems for Energy Reduction: A Graph Neural Network Approach,

    T. Feys et al., “Learning to Quantize and Precode in Massive MIMO Systems for Energy Reduction: A Graph Neural Network Approach,” IEEE J. Sel. Topics Signal Process., 2025

    work page 2025

  11. [11]

    The Z3RO Family of Precoders Cancelling Nonlinear Power Amplification Distortion in Large Array Systems,

    F. Rottenberg et al., “The Z3RO Family of Precoders Cancelling Nonlinear Power Amplification Distortion in Large Array Systems,” IEEE Trans. Wireless Commun., Mar. 2023

    work page 2023

  12. [12]

    Leveraging Power Amplifier Distortion for Physical Layer Security,

    R. G. Alavicheh et al., “Leveraging Power Amplifier Distortion for Physical Layer Security,” inPIMRC, Sep. 2025

    work page 2025

  13. [13]

    Optimal design of energy-efficient multi-user MIMO systems: Is massive MIMO the answer?

    E. Björnson et al., “Optimal design of energy-efficient multi-user MIMO systems: Is massive MIMO the answer?”IEEE Trans. Wire- less Commun., vol. 14, no. 6, 2015

    work page 2015

  14. [14]

    Techtile – Open 6G R&D Testbed for Com- munication, Positioning, Sensing, WPT and Federated Learning,

    G. Callebaut et al., “Techtile – Open 6G R&D Testbed for Com- munication, Positioning, Sensing, WPT and Federated Learning,” in EuCNC/6G Summit, Jun. 2022

    work page 2022

  15. [15]

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

    T. Miao et al., “GNN-based Precoder Design and Fine-tuning for Cell-free Massive MIMO with Real-world CSI,” inSITB, May 2025

    work page 2025

  16. [16]

    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)

    work page 2022

  17. [17]

    Adam: A method for stochastic optimization,

    D. P. Kingma et al., “Adam: A method for stochastic optimization,” inICLR, 2014

    work page 2014

  18. [18]

    Single Versus Multi-Tone Wireless Power Transfer with Physically Large Arrays,

    J. Van Mulders et al., “Single Versus Multi-Tone Wireless Power Transfer with Physically Large Arrays,” in6GNet, Oct. 2024

    work page 2024