Pith. sign in

REVIEW 4 major objections 5 minor 40 references

CaFTRA claims that a Transformer can map user geolocation to per-resource-block MIMO parameters, eliminating real-time CSI feedback and outperforming 5G feedback-based transmission under mobility.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 18:44 UTC pith:RL4Q6AHM

load-bearing objection The core claim—per-RB CSI predictable from geolocation—is only tested in one static channel drop, and the matching proofs are not sound; the architecture is plausible but the paper needs major revision. the 4 major comments →

arxiv 2512.03767 v3 pith:RL4Q6AHM submitted 2025-12-03 eess.SY cs.ITcs.SYmath.IT

CaFTRA: Frequency-Domain Correlation-Aware Feedback-Free MIMO Transmission and Resource Allocation for 6G and Beyond

Bo Qian , Hanlin Wu , Jiacheng Chen , Yunting Xu , Xiaoyu Wang , Haibo Zhou , Yusheng Ji This is my paper
classification eess.SY cs.ITcs.SYmath.IT
keywords feedback-free MIMO transmissionCSI prediction from user geolocationlearnable query embeddingsTransformer networkfrequency-domain correlationresource allocationmany-to-one matchingfully-decoupled radio access networks
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper proposes a framework in which downlink MIMO transmission is driven entirely by predicted channel state information derived from user geolocation, with no real-time uplink feedback. The central claim is that at 3.5 GHz the per-resource-block parameters that standard MIMO feedback reports — rank indicator, two channel-quality indicators, and precoding matrix indicator — can be learned as a function of where the user is located. To do this, a Transformer network with learnable query embeddings is trained to exploit frequency-domain correlations across resource blocks, and a many-to-one matching algorithm allocates multiple base stations and resource blocks to users under coverage-expanded downlink. The payoff appears in the mobility results: where a 3 ms feedback loop degrades 5G, the geolocation-based prediction gains roughly 20% throughput at 10 km/h and 93% at 15 km/h over 5G closed-loop spatial multiplexing.

Core claim

The paper's central claim is that MIMO CSI feedback can be replaced wholesale by a learned map from user location to CSI parameters. The proposed LQTN predicts, for every resource block and every base station, the full set of parameters a standard-compliant receiver would otherwise feed back — RI, CQI1, CQI2, and PMI — given only the base station and user geolocations. Because the mapping is learned per frequency resource with attention across resource blocks, it captures how neighboring resource blocks behave together rather than treating each independently. With the parameters in hand, the transmitter selects modulation and coding, number of spatial layers, and precoding matrix as it would

What carries the argument

The Learnable Queries-driven Transformer Network (LQTN): an encoder-decoder Transformer in which each resource block is assigned a trainable query embedding; multi-head attention lets the decoder aggregate spatial context from base-station and user positions while also modeling correlations across resource blocks. The same network outputs RI, CQI1, CQI2, and PMI for all RBs at once. The second component is the M3-MAMA matching algorithm: it binds base-station and resource-block pairs and formulates the NP-hard 0-1 integer program as a many-to-one stable matching problem, repeatedly attempting swaps and reassignments with O(W^2) complexity until no improving exchange exists, with a proof that

Load-bearing premise

The paper's results rest on the premise that per-resource-block RI, CQI, and PMI are a deterministic, smooth function of user geolocation at 3.5 GHz; real multipath channels vary on a scale of centimeters whereas geolocation is meter-scale, so if the mapping is not smooth or stationary, predicted parameters are wrong exactly when feedback delay hurts 5G and the mobility gains vanish.

What would settle it

Take a trained CSI map, deploy it on a route the model has never seen, and simultaneously measure ground-truth CSI via standard feedback while the scattering environment changes (a vehicle passes, foliage changes, or a building is altered). If predicted PMI and CQI diverge from ground truth at locations whose neighbors were included in training, at speeds where a 3 ms feedback loop would still be fresh, the core claim fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Feedback-free MIMO removes the uplink feedback channel and its delay from the transmission loop; downlink coverage in FD-RAN can expand because the downlink base station no longer depends on the user's uplink capability.
  • In high mobility, geolocation-based prediction beats feedback: the paper reports about 20% per-user throughput gain over 5G closed-loop spatial multiplexing at 10 km/h and about 93% at 15 km/h, at a static penalty of less than 14%.
  • The frequency-domain correlation-aware model lowers prediction error relative to independent per-RB prediction, while the shared-query architecture uses about 84% fewer parameters than the independent baseline.
  • The matching allocator, proven to converge to a pairwise-stable matching, lifts spectral efficiency by roughly 60% over round-robin and 15% over best-CQI scheduling, while also improving fairness under minimum-resource constraints.
  • The training loop depends on collecting historical (geolocation, CSI) labels; the paper proposes periodic feedback and retraining to keep predictions current as the environment changes.

Where Pith is reading between the lines

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

  • The stated gains implicitly assume the geolocation-to-CSI map is stationary across time and generalize to unmeasured locations; real deployments with moving scatterers, weather shifts, or changed building geometry would likely require frequent retraining, and the paper does not model positioning error.
  • Because the heatmaps show CSI varying smoothly with location, the model may interpolate well between sampled points but could fail on extrapolated regions; a route-based evaluation with held-out trajectories would directly test this boundary.
  • The 93% mobility gain is measured against a feedback loop with 3 ms delay; practical feedback delays often exceed that, so the real-world gap could be larger, but the scheme's own training feedback creates a residual dependence on feedback that should be factored into net overhead.
  • The same frequency-correlation-aware structure might extend to other codebooks or frequency bands, though the granularity of PMI and the smoothness of the mapping would need fresh validation at each new carrier frequency or scattering environment.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes CaFTRA, a framework for fully-decoupled RAN (FD-RAN) in 6G, with two main components. At the PHY layer, a Learnable Queries-driven Transformer Network (LQTN) predicts per-RB CSI parameters (RI, CQI1, CQI2, PMI) from BS and UE geolocations, aiming to eliminate real-time CSI feedback. At the MAC layer, a many-to-one matching algorithm (M3-MAMA) is used for multi-BS association and multi-RB allocation, with claims of pairwise-stable convergence. Simulations in the Vienna 5G system-level simulator compare the proposal with 5G CLSM, Round-Robin, and Best CQI schedulers, reporting throughput, spectral-efficiency, and fairness gains, including a 93% per-user throughput gain over CLSM at 15 km/h. The held-out geolocation test set is a genuine out-of-sample spatial interpolation test, but the evaluation is confined to a single static channel drop and the matching proof is not valid as written.

Significance. If the geolocation-to-CSI mapping were validated across independent channel realizations, the framework would be a significant step toward feedback-free MIMO in FD-RAN, with plausible value for AI-native 6G. The LQTN architecture is a reasonable, parameter-sharing design, and the paper honestly reports an ablation against an independent Transformer baseline, along with complexity numbers. The use of a system-level simulator and out-of-sample geolocations is a strength. However, the central physical premise — that instantaneous PMI/CQI is a deterministically smooth function of geolocation — is only demonstrated within one static drop, and the stability proof for the MAC algorithm is internally incoherent. Therefore the reported gains are not yet established as generalizable.

major comments (4)
  1. [§VI.B.1, Figs. 6, 8, 9] The load-bearing claim is that per-RB RI/CQI1/CQI2/PMI can be predicted from geolocation alone, eliminating real-time feedback. The evidence for this is the heatmaps in Fig. 6 and the held-out-location test, but all data come from a single static drop of the Vienna simulator. Within one drop the small-scale fading realization is fixed, so CSI is by construction a deterministic function of position; the 5,000 held-out locations test spatial interpolation within that same realization, not generalization across independent fading states. At 3.5 GHz the wavelength is ~8.6 cm and PMI/CQI depend on the instantaneous channel matrix, so at a fixed location they will vary from one coherence interval to the next in a real environment. The paper also concedes non-stationarity in §VI.B.1 by requiring periodic CSI feedback and retraining, which conflicts with the abstract's 'eliminates real-time upli
  2. [§V.B, Definition 2, Lemma 1, Theorem 2, Algorithm 1] The proof of Lemma 1 is not a valid proof. The symbols φ and the preference relation ≻ are not defined, and the argument is logically incoherent: it asserts stability from the rejection history of a myopic greedy exchange heuristic without any formal connection to the stated Definition 2. Theorem 2's proof states that 'the algorithm only accepts allocations or exchanges that strictly improve the total system throughput', but Algorithm 1 explicitly selects the highest of T0, T1, T2, T3 and can leave the assignment unchanged (T0 is allowed). Thus the claimed guarantee of convergence to a pairwise-stable matching is unproven for the algorithm as written. Please correct the definitions/proof, or remove the formal stability claim and state that M3-MAMA is a throughput-improving heuristic whose convergence is only demonstrated empirically.
  3. [§VI.B.2, Fig. 9] The headline mobility result (≈20% gain at 10 km/h and ≈93% at 15 km/h over 5G CLSM with 3 ms feedback delay) is not given a credible mechanism. At 15 km/h and 3.5 GHz, a UE moves only ~1.25 cm in 3 ms (about 0.15λ), which is much smaller than the typical channel decorrelation distance; the corresponding coherence time is on the order of 10 ms. A 93% degradation of CLSM under this delay is surprising and may be an artifact of the simulator's feedback period, HARQ model, or CQI quantization rather than of the geolocation-prediction approach. Please report the CLSM feedback period, the actual CSI age, the channel model parameters, and a sensitivity study over feedback delay and velocity, so the Fig. 9 gain can be interpreted.
  4. [§VI.B.3, Figs. 10–13] The MAC-layer comparisons evaluate CaFTRA in an FD-RAN setting where a UE can be served by RBs from multiple BSs, against 5G schedulers that are restricted to a single BS (single-connectivity). The 15–60% spectral-efficiency/fairness gains may therefore be attributable to the FD-RAN architecture (larger pool of assignable RBs) rather than to the proposed feedback-free prediction or the M3-MAMA algorithm. To support the specific claims about the proposed resource-allocation algorithm, please include a same-connectivity baseline or an ablation that decomposes the gain into architecture-driven and algorithm-driven components.
minor comments (5)
  1. [§IV heading and §VI.B.1] Typos and inconsistent notation: 'among Bs' should be 'among RBs'; 'extented', 'geoglocations', and 'Jain's Fairnes Index' appear in the text. In §VI.B.1, the statement that users 'periodically feedback CSI' should be reconciled with the claimed absence of CSI feedback.
  2. [§V.B, Definition 3] The notation W_w(\hat{M}) and the choice-set operator are not formally introduced. Also, in Definition 2, since μ(w) is a single UE, the condition should be written m ≠ μ(w) rather than m ∉ μ(w).
  3. [Table I] The parameter counts should be explained: the Independent Transformer Network uses a reduced embedding dimension (256) yet reports 275.7M parameters, while CaFTRA uses 1024 dimensions and reports 44.0M. If the baseline duplicates weights across 100 per-RB networks, state this explicitly.
  4. [Fig. 8] The y-axis label is garbled in the render (appears as 'Normalized Mean Absolute Error' in corrupted encoding). Please regenerate the figure with a clean label.
  5. [§V.A] The rate calculation example is for CQI=1 and yields 0.020064 Mbps per RB; it would be helpful to state that this is the per-RB rate and to note the implied peak rate at high CQI so readers can reconcile with the ~100 Mbps throughputs in later figures.

Circularity Check

0 steps flagged

No significant circularity: the CSI-prediction result is an out-of-sample generalization test and the scheduler claims do not reduce to the optimized objective.

full rationale

The paper's derivation chain is not circular by construction. The LQTN is trained on 50,000 simulator-generated (geolocation, CSI) samples and tested on a separate set of 5,000 randomly generated geolocations ('Another 5,000 randomly generated user geolocations are used as test data to evaluate the CSI prediction accuracy'), so the reported CSI-prediction performance is a genuine out-of-sample generalization result rather than a refit of the training set. The MAC-layer objective in Eq. (9) maximizes Rate[CSI_LQTN], but the reported spectral-efficiency and fairness comparisons are implemented in the Vienna 5G system-level simulator, an external simulator, so the scheduler is not evaluated solely by the same estimated-rate function it optimizes. The matching convergence claims (Lemma 1 and Theorem 2) are independent of the learned model and do not import a uniqueness or fit result from prior work. The related-work section does cite the authors' own FD-RAN and feedback-free papers ([14], [30], [31]), and the 'feedback-free' phrasing is qualified in Section VI.B.1 by periodic CSI feedback and retraining, but these are novelty/scoping statements rather than load-bearing derivations. The main correctness risks are the stationarity of the geolocation-to-CSI map across independent channel realizations and the under-developed proof of Lemma 1; both are validity/rigor concerns, not circular reductions. Therefore the circularity score is 0.

Axiom & Free-Parameter Ledger

4 free parameters · 6 axioms · 0 invented entities

The central claim rests on three external inputs the reader must take on faith: (1) the simulator's deterministic position-to-CSI map, (2) a label-collection channel that partially contradicts the feedback-free framing, and (3) an unproven property of the exchange algorithm. The only hand-chosen numbers are ML hyperparameters (embedding dims, query count) and the 3 ms comparator delay. No new physical entities are introduced; the BS-RB pair is a notational device, and the learnable queries are ordinary trainable parameters.

free parameters (4)
  • LQTN embedding dimension = 1024
    Hand-chosen input embedding size for geolocation features; no sweep or sensitivity analysis, and it sets the capacity of the central CSI predictor.
  • Independent-Transformer baseline embedding dimension = 256
    The ablation baseline deliberately uses a 256-dim embedding vs CaFTRA's 1024, confounding the frequency-correlation comparison with model capacity (Table I, Fig. 8).
  • Feedback delay for 5G CLSM comparator = 3 ms
    Paper states this is an optimistic assumption for CLSM; CaFTRA is also assigned the same 3 ms delay despite being feedback-free, which the text never explains (Section VI.B.2).
  • Number of learnable queries = 100 per BS (one per RB)
    One learnable query per RB by construction; natural but hand-chosen without studying sensitivity to query count.
axioms (6)
  • domain assumption CSI (RI, CQI, PMI) is a deterministic smooth function of geolocation at RB granularity
    Invoked in Section IV.A and justified by Fig. 6 heatmaps from a static Vienna drop; without it, per-RB PMI/CQI prediction from meter-scale position fails in real channels.
  • domain assumption Exact user geolocation is available without error
    The evaluation feeds simulator-exact coordinates to LQTN; no GPS/positioning-error model is included, while real error is meters — above the cm-scale fading coherence distance.
  • domain assumption Interference between DL BSs is negligible because spectrum is partitioned
    Section V.B: the whole spectrum band is equally divided to x BSs and they do not share spectrum to avoid interference; this removes the multi-cell interference that FD-RAN coordination must handle in practice.
  • domain assumption Rate(CSI) with accurate PMI gives achievable throughput
    Section V.A: provided the PMI is accurately predicted, the frame error rate remains low; the scheduler optimizes this estimated rate, but predicted-PMI error enters actual throughput only in the later Vienna stage.
  • ad hoc to paper Greedy pairwise exchange yields a pairwise-stable matching
    Lemma 1 proof (Section V.B) asserts stability without a valid argument; substitutability (Definition 3) is never used. This is the unproven bridge from the algorithm to the stability claim.
  • domain assumption Historical (geolocation, CSI) labels are available via periodic feedback
    Section VI.B.1: users periodically feedback CSI to the control BS, aggregated by the edge cloud as training data; feedback-free operation still presupposes a feedback channel for labels.

pith-pipeline@v1.3.0-alltime-deepseek · 5211 in / 5942 out tokens · 283085 ms · 2026-08-03T18:44:26.237784+00:00 · methodology

0 comments
read the original abstract

The fundamental designs of wireless systems toward AI-Native 6G and beyond are driven by the need for ever-increasing demand of mobile data traffic, extreme spectral efficiency, and adaptability across diverse service scenarios. To overcome the limitations posed by feedback-based multiple-input and multiple-output (MIMO) transmission, we propose a novel frequency-domain Correlation-aware Feedback-free MIMO Transmission and Resource Allocation (CaFTRA) framework tailored for fully-decoupled radio access networks (FD-RAN) to meet the emerging requirements of AI-Native 6G and beyond. By leveraging artificial intelligence (AI), CaFTRA effectively eliminates real-time uplink feedback by predicting channel state information (CSI) based solely on user geolocation. We introduce a Learnable Queries-driven Transformer Network for CSI mapping from user geolocation, which utilizes multi-head attention and learnable query embeddings to accurately capture frequency-domain correlations among resource blocks (RBs), thereby significantly improving the precision of CSI prediction. Once base stations (BSs) adopt feedback-free transmission, their downlink transmission coverage can be significantly expanded due to the elimination of frequent uplink feedback. To enable efficient resource scheduling under such extensive-coverage scenarios, we apply a low-complexity many-to-one matching theory-based algorithm for efficient multi-BS association and multi-RB resource allocation, which is proven to converge to a stable matching within limited iterations. Simulation results demonstrate that CaFTRA achieves stable matching convergence and significant gains in spectral efficiency and user fairness compared to 5G, underscoring its potential value for 6G standardization efforts.

Figures

Figures reproduced from arXiv: 2512.03767 by Bo Qian, Haibo Zhou, Hanlin Wu, Jiacheng Chen, Xiaoyu Wang, Yunting Xu, Yusheng Ji.

Figure 1
Figure 1. Figure 1: The Proposed CaFTRA Framework in FD-RAN. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Working Flow of Frequency-Domain Correlation-Aware Feedback [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Schematic Diagram of the LQTN-based CSI Prediction Model. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Schematic Diagram of the Multi-Head Attention Mechanism. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: 2D (left) and 3D (right) Views of the Peng Cheng Laboratory Scenario [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Heatmap of historical CSI in training dataset. [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 8
Figure 8. Figure 8: Normalized Mean Absolute Error of CSI Prediction Comparison [PITH_FULL_IMAGE:figures/full_fig_p010_8.png] view at source ↗
Figure 7
Figure 7. Figure 7: Throughput (Mbps) Heatmap Comparision of 5G CLSM and CaFTRA [PITH_FULL_IMAGE:figures/full_fig_p010_7.png] view at source ↗
Figure 9
Figure 9. Figure 9: Per-User Throughput Comparison for 5G CLSM and CaFTRA-based [PITH_FULL_IMAGE:figures/full_fig_p010_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Spectral Efficiency Comparison of 5G Round-Robin, 5G Best CQI, [PITH_FULL_IMAGE:figures/full_fig_p011_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Jain’s Fairness Index Comparison of 5G Best CQI and CaFTRA [PITH_FULL_IMAGE:figures/full_fig_p011_11.png] view at source ↗
Figure 13
Figure 13. Figure 13: CDF of Per-RB Throughput for 5G Best CQI and CaFTRA-based [PITH_FULL_IMAGE:figures/full_fig_p012_13.png] view at source ↗
Figure 12
Figure 12. Figure 12: Average Per-User Throughput Comparison of 5G Round-Robin, 5G [PITH_FULL_IMAGE:figures/full_fig_p012_12.png] view at source ↗
Figure 14
Figure 14. Figure 14: Convergence of Exchange-Matching Iterations in M3-MAMA. [PITH_FULL_IMAGE:figures/full_fig_p013_14.png] view at source ↗

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

40 extracted references

  1. [1]

    Towards Sustainability in 6G and beyond: Challenges and Opportunities of Open RAN,

    H. Ahmadi, M. Rahmani, S. B. Chetty, E. E. Tsiropoulou, H. Arslan, M. Debbah, and T. Quek, “Towards Sustainability in 6G and beyond: Challenges and Opportunities of Open RAN,”IEEE Communications Standards Magazine, vol. 9, no. 3, pp. 126–135, 2025

    2025

  2. [2]

    5G-Advanced Toward 6G: Past, Present, and Future,

    W. Chen, X. Lin, J. Lee, A. Toskala, S. Sun, C. F. Chiasserini, and L. Liu, “5G-Advanced Toward 6G: Past, Present, and Future,”IEEE Journal on Selected Areas in Communications, vol. 41, no. 6, pp. 1592–1619, 2023

    2023

  3. [3]

    Toward User-Centric Resource Allocation for 6G: An Economic Perspective,

    J. Chen, B. Qian, Y . Xu, H. Zhou, and X. Shen, “Toward User-Centric Resource Allocation for 6G: An Economic Perspective,”IEEE Network, vol. 37, no. 2, pp. 254–261, 2023

    2023

  4. [4]

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

    C.-X. Wang, X. You, X. Gao, X. Zhu, Z. Li, C. Zhang, H. Wang, Y . Huang, Y . Chen, H. Haas, J. S. Thompson, E. G. Larsson, M. D. Renzo, W. Tong, P. Zhu, X. Shen, H. V . Poor, and L. Hanzo, “On the Road to 6G: Visions, Requirements, Key Technologies, and Testbeds,” IEEE Communications Surveys & Tutorials, vol. 25, no. 2, pp. 905–974, 2023

    2023

  5. [5]

    Multi-Operator Spectrum Sharing for Massive IoT Coexisting in 5G/B5G Wireless Networks,

    B. Qian, H. Zhou, T. Ma, K. Yu, Q. Yu, and X. Shen, “Multi-Operator Spectrum Sharing for Massive IoT Coexisting in 5G/B5G Wireless Networks,”IEEE Journal on Selected Areas in Communications, vol. 39, no. 3, pp. 881–895, 2021

    2021

  6. [6]

    3C Resource Sharing for Personalized Content Delivery in B5G Networks: A Contract Approach,

    B. Qian, T. Ma, K. Yu, Y . Xu, Y . Wu, and H. Zhou, “3C Resource Sharing for Personalized Content Delivery in B5G Networks: A Contract Approach,”IEEE Internet of Things Journal, vol. 10, no. 15, pp. 13 442– 13 457, 2023

    2023

  7. [7]

    AI-Assisted Network-Slicing Based Next-Generation Wireless Networks,

    X. Shen, J. Gao, W. Wu, K. Lyu, M. Li, W. Zhuang, X. Li, and J. Rao, “AI-Assisted Network-Slicing Based Next-Generation Wireless Networks,”IEEE Open Journal of V ehicular Technology, vol. 1, pp. 45–66, 2020

    2020

  8. [8]

    Lever- aging Dynamic Stackelberg Pricing Game for Multi-Mode Spectrum Sharing in 5G-V ANET,

    B. Qian, H. Zhou, T. Ma, Y . Xu, K. Yu, X. Shen, and F. Hou, “Lever- aging Dynamic Stackelberg Pricing Game for Multi-Mode Spectrum Sharing in 5G-V ANET,”IEEE Transactions on V ehicular Technology, vol. 69, no. 6, pp. 6374–6387, 2020

    2020

  9. [9]

    Digital Twin of Channel: Diffusion Model for Sensing-Assisted Statistical Channel State Information Generation,

    X. Gong, X. Liu, A.-A. Lu, X. Gao, X.-G. Xia, C.-X. Wang, and X. You, “Digital Twin of Channel: Diffusion Model for Sensing-Assisted Statistical Channel State Information Generation,”IEEE Transactions on Wireless Communications, vol. 24, no. 5, pp. 3805–3821, 2025

    2025

  10. [10]

    Deep CSI Compression for Dual-Polarized Massive MIMO Channels With Disentangled Representation Learning,

    S. Fan, W. Xu, R. Xie, S. Jin, D. W. K. Ng, and N. Al-Dhahir, “Deep CSI Compression for Dual-Polarized Massive MIMO Channels With Disentangled Representation Learning,”IEEE Transactions on Communications, vol. 72, no. 9, pp. 5564–5580, 2024

    2024

  11. [11]

    ArguteDUB: Deep Learning Based Distributed Uplink Beamforming in 6G-Based IoV,

    X. Yi, J. Li, Y . Liu, L. Kong, Y . Shao, G. Chen, X. Liu, S. Mumtaz, and J. J. P. C. Rodrigues, “ArguteDUB: Deep Learning Based Distributed Uplink Beamforming in 6G-Based IoV,”IEEE Transactions on Mobile Computing, vol. 23, no. 4, pp. 2551–2565, 2024

    2024

  12. [12]

    Near-Field Beam Training for Extremely Large-Scale MIMO Based on Deep Learning,

    J. Nie, Y . Cui, Z. Yang, W. Yuan, and X. Jing, “Near-Field Beam Training for Extremely Large-Scale MIMO Based on Deep Learning,” IEEE Transactions on Mobile Computing, vol. 24, no. 1, pp. 352–362, 2025

    2025

  13. [13]

    Foundations of User- Centric Cell-Free Massive MIMO,

    ¨O. T. Demir, E. Bj ¨ornson, and L. Sanguinetti, “Foundations of User- Centric Cell-Free Massive MIMO,”F oundations and Trends®in Signal Processing, vol. 14, no. 3-4, pp. 162–472, 2021

    2021

  14. [14]

    A Fully-Decoupled RAN Architecture for 6G Inspired by Neurotrans- mission,

    Q. Yu, H. Zhou, J. Chen, Y . Li, J. Jing, J. Zhao, B. Qian, and J. Wang, “A Fully-Decoupled RAN Architecture for 6G Inspired by Neurotrans- mission,”Journal of Communications and Information Networks, vol. 4, no. 4, pp. 15–23, 2019

    2019

  15. [15]

    Evolution of RAN Architectures Toward 6G: Motivation, Development, and Enabling Technologies,

    J. Chen, X. Liang, J. Xue, Y . Sun, H. Zhou, and X. Shen, “Evolution of RAN Architectures Toward 6G: Motivation, Development, and Enabling Technologies,”IEEE Communications Surveys & Tutorials, vol. 26, no. 3, pp. 1950–1988, 2024

    1950

  16. [16]

    Why to decouple the uplink and downlink in cellular networks and how to do it,

    F. Boccardi, J. Andrews, H. Elshaer, M. Dohler, S. Parkvall, P. Popovski, and S. Singh, “Why to decouple the uplink and downlink in cellular networks and how to do it,”IEEE Communications Magazine, vol. 54, no. 3, pp. 110–117, 2016

    2016

  17. [17]

    Fully- Decoupled Radio Access Networks: A Resilient Uplink Base Stations Cooperative Reception Framework,

    J. Zhao, Q. Yu, B. Qian, K. Yu, Y . Xu, H. Zhou, and X. Shen, “Fully- Decoupled Radio Access Networks: A Resilient Uplink Base Stations Cooperative Reception Framework,”IEEE Transactions on Wireless Communications, vol. 22, no. 8, pp. 5096–5110, 2023

    2023

  18. [18]

    Attention Is All You Need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention Is All You Need,”Advances in Neural Information Processing Systems, vol. 30, pp. 5998–6008, 2017

    2017

  19. [19]

    End-to-End Object Detection With Transformers,

    N. Carion, F. Massa, G. Synnaeve, N. Usunier, A. Kirillov, and S. Zagoruyko, “End-to-End Object Detection With Transformers,” in European Conference on Computer Vision. Springer, 2020, pp. 213– 229. 14

    2020

  20. [20]

    Flexible Multi-Node Simulation of Cellular Mobile Communications: the Vienna 5G System Level Simulator,

    M. K. M ¨uller, F. Ademaj, T. Dittrich, A. Fastenbauer, B. R. Elbal, A. Nabavi, L. Nagel, S. Schwarz, and M. Rupp, “Flexible Multi-Node Simulation of Cellular Mobile Communications: the Vienna 5G System Level Simulator,”EURASIP Journal on Wireless Communications and Networking, vol. 2018, no. 1, p. 227, 2018

    2018

  21. [21]

    Overview of Deep Learning- Based CSI Feedback in Massive MIMO Systems,

    J. Guo, C.-K. Wen, S. Jin, and G. Y . Li, “Overview of Deep Learning- Based CSI Feedback in Massive MIMO Systems,”IEEE Transactions on Communications, vol. 70, no. 12, pp. 8017–8045, 2022

    2022

  22. [22]

    The Standardization of 5G-Advanced in 3GPP,

    W. Chen, J. Montojo, J. Lee, M. Shafi, and Y . Kim, “The Standardization of 5G-Advanced in 3GPP,”IEEE Communications Magazine, vol. 60, no. 11, pp. 98–104, 2022

    2022

  23. [23]

    Deep Learning for CSI Feedback: One-Sided Model and Joint Multi-Module Learning Perspectives,

    Y . Guo, W. Chen, F. Sun, J. Cheng, M. Matthaiou, and B. Ai, “Deep Learning for CSI Feedback: One-Sided Model and Joint Multi-Module Learning Perspectives,”IEEE Communications Magazine, vol. 63, no. 7, pp. 90–97, 2025

    2025

  24. [24]

    Transformer Network Based Channel Prediction for CSI Feedback En- hancement in AI-Native Air Interface,

    T. Zhou, X. Liu, Z. Xiang, H. Zhang, B. Ai, L. Liu, and X. Jing, “Transformer Network Based Channel Prediction for CSI Feedback En- hancement in AI-Native Air Interface,”IEEE Transactions on Wireless Communications, vol. 23, no. 9, pp. 11 154–11 167, 2024

    2024

  25. [25]

    Predict- ing Wireless Channel Features Using Neural Networks,

    S. Navabi, C. Wang, O. Y . Bursalioglu, and H. Papadopoulos, “Predict- ing Wireless Channel Features Using Neural Networks,” in2018 IEEE International Conference on Communications (ICC), 2018, pp. 1–6

    2018

  26. [26]

    Feature Extraction Using Hough Transform in Radio Propagation Estimation,

    R. Hagiwara, K. Ichige, T. Nagao, and T. Hayashi, “Feature Extraction Using Hough Transform in Radio Propagation Estimation,” in2023 IEEE 98th V ehicular Technology Conference (VTC2023-Fall), 2023, pp. 1–5

    2023

  27. [27]

    Fully-Decoupled Radio Access Networks: A Flexible Downlink Multi- connectivity and Dynamic Resource Cooperation Framework,

    K. Yu, Q. Yu, Z. Tang, J. Zhao, B. Qian, Y . Xu, H. Zhou, and X. Shen, “Fully-Decoupled Radio Access Networks: A Flexible Downlink Multi- connectivity and Dynamic Resource Cooperation Framework,”IEEE Transactions on Wireless Communications, vol. 22, no. 6, pp. 4202– 4214, 2023

    2023

  28. [28]

    Enabling Fully-Decoupled Radio Access With Elastic Resource Allocation,

    B. Qian, T. Ma, Y . Xu, J. Zhao, K. Yu, Y . Wu, and H. Zhou, “Enabling Fully-Decoupled Radio Access With Elastic Resource Allocation,”IEEE Transactions on Cognitive Communications and Networking, vol. 9, no. 4, pp. 1025–1040, 2023

    2023

  29. [29]

    Federated Learn- ing Over Fully-Decoupled RAN Architecture for Two-Tier Computing Acceleration,

    Y . Xu, B. Qian, K. Yu, T. Ma, L. Zhao, and H. Zhou, “Federated Learn- ing Over Fully-Decoupled RAN Architecture for Two-Tier Computing Acceleration,”IEEE Journal on Selected Areas in Communications, vol. 41, no. 3, pp. 789–801, 2023

    2023

  30. [30]

    Enabling Feedback-Free MIMO Transmission for FD-RAN: A Data-driven Approach,

    J. Liu, J. Chen, Z. Liu, and H. Zhou, “Enabling Feedback-Free MIMO Transmission for FD-RAN: A Data-driven Approach,”IEEE Transac- tions on Mobile Computing, vol. 24, no. 3, pp. 2437–2454, 2025

    2025

  31. [31]

    Fully-Decoupled RAN for Feedback-Free Multi-Base Station Trans- mission in MIMO-OFDM System,

    Y . Xu, Z. Liu, B. Qian, H. Du, J. Chen, J. Kang, H. Zhou, and D. Niyato, “Fully-Decoupled RAN for Feedback-Free Multi-Base Station Trans- mission in MIMO-OFDM System,”IEEE Journal on Selected Areas in Communications, vol. 43, no. 3, pp. 780–794, 2025

    2025

  32. [32]

    NR; Physical layer procedures for data,

    3GPP, “NR; Physical layer procedures for data,” 3rd Generation Partnership Project (3GPP), TS 38.214, 2024, V18.4.0, Release 18

    2024

  33. [33]

    Calculation of the Spatial Preprocessing and Link Adaption Feedback for 3GPP UMTS/LTE,

    S. Schwarz, C. Mehlf ¨uhrer, and M. Rupp, “Calculation of the Spatial Preprocessing and Link Adaption Feedback for 3GPP UMTS/LTE,” in 2010 Wireless Advanced 2010, 2010, pp. 1–6

    2010

  34. [34]

    Link performance models for system level simulations of broadband radio access systems,

    K. Brueninghaus, D. Astely, T. Salzer, S. Visuri, A. Alexiou, S. Karger, and G.-A. Seraji, “Link performance models for system level simulations of broadband radio access systems,” in2005 IEEE 16th International Symposium on Personal, Indoor and Mobile Radio Communications, vol. 4, 2005, pp. 2306–2311

    2005

  35. [35]

    Capacity of Bit-Interleaved Channels,

    G. Caire, G. Taricco, and E. Biglieri, “Capacity of Bit-Interleaved Channels,”Electronics Letters, vol. 32, no. 12, pp. 1060–1061, 1996

    1996

  36. [36]

    MIMO Beamforming Using PMI Type II Precoding,

    R. M. P. P. Sudhakar and C. Jenisha, “MIMO Beamforming Using PMI Type II Precoding,”KTH ROYAL INSTITUTE OF TECHNOLOGY, 2021

    2021

  37. [37]

    Enhanced Uplink Resource Allocation in Non-Orthogonal Multiple Access Systems,

    R. Ruby, S. Zhong, H. Yang, and K. Wu, “Enhanced Uplink Resource Allocation in Non-Orthogonal Multiple Access Systems,”IEEE Trans- actions on Wireless Communications, vol. 17, no. 3, pp. 1432–1444, 2018

    2018

  38. [38]

    Joint MCS Adaptation and RB Allocation in Cellular Networks Based on Deep Reinforcement Learning With Stable Matching,

    X. Ye and L. Fu, “Joint MCS Adaptation and RB Allocation in Cellular Networks Based on Deep Reinforcement Learning With Stable Matching,”IEEE Transactions on Mobile Computing, vol. 23, no. 1, pp. 549–565, 2024

    2024

  39. [39]

    A Quantitative Measure of Fairness and Discrimination,

    R. K. Jain, D.-M. W. Chiu, W. R. Haweet al., “A Quantitative Measure of Fairness and Discrimination,”Eastern Research Laboratory, Digital Equipment Corporation, Hudson, MA, vol. 21, pp. 1–38, 1984

    1984

  40. [40]

    Toward Collision- Free and Efficient Coordination for Automated Vehicles at Unsignalized Intersection,

    B. Qian, H. Zhou, F. Lyu, J. Li, T. Ma, and F. Hou, “Toward Collision- Free and Efficient Coordination for Automated Vehicles at Unsignalized Intersection,”IEEE Internet of Things Journal, vol. 6, no. 6, pp. 10 408– 10 420, 2019

    2019