pith. sign in

arxiv: 2605.16141 · v1 · pith:XVXGKB43new · submitted 2026-05-15 · 📡 eess.SP

SiFo: Wireless Foundation Model for Low-Overhead Site-Specific CSI Feedback

Cheng-Jie Zhao , Zhaolin Wang , Zongyao Zhao , Yuanwei Liu This is my paper

Pith reviewed 2026-05-19 21:42 UTC · model grok-4.3

classification 📡 eess.SP
keywords CSI feedbacksite-specificfoundation modelRSRPlow-overhead3GPP NR Type-IIspectral efficiencywireless adaptation
0
0 comments X

The pith

SiFo pretrains a CSI feedback model across sites and adapts it to new deployments by matching users to calibration samples via RSRP measurements without parameter updates.

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

SiFo addresses the scalability limit of site-specific CSI feedback by pretraining a model on multiple source sites to learn shared propagation patterns instead of training a fresh network for every new deployment. At a target site a small calibration set of users reports RSRP fingerprints during standard SSB probing and stores the corresponding full-CSI subspace labels. In online operation a served user is matched to the nearest calibration samples through the identical low-overhead RSRP procedure, so the model supplies site-specific subspace guidance without any retraining. This yields higher CSI-capture efficiency than separately trained per-site baselines under the same labeled-data budget and approaches the accuracy of full 3GPP NR Type-II feedback while using far less overhead. The resulting efficiency gains produce measurable spectral-efficiency improvements when target-site data remain limited.

Core claim

SiFo pretrains a CSI feedback model across source sites to capture common propagation structures and adapts it to a target site through lightweight calibration. A small set of target-site users reports low-dimensional RSRP fingerprints during SSB probing and their full-CSI subspace labels are stored as calibration memory. During online operation a served user is matched to nearby calibration samples via the same SSB probing and RSRP reporting, allowing the model to deliver site-specific subspace guidance without updating any parameters. This transfers common knowledge while retaining local adaptation and produces higher CSI-capture efficiency than per-site baselines under identical target-s.

What carries the argument

RSRP-based user matching to a stored calibration memory that pairs low-dimensional reference-signal fingerprints with full-CSI subspace labels, supplying site-specific guidance without model-parameter updates.

If this is right

  • SiFo achieves higher CSI-capture efficiency than separately trained site-specific learning baselines under the same target-site labeled budget.
  • SiFo approaches the performance of high-overhead 3GPP NR Type-II feedback using only RSRP measurements collected during online SSB probing.
  • The high CSI-capture efficiency and low overhead convert into effective spectral efficiency improvement under limited target-site data.
  • SiFo scales to many deployments by avoiding the need to train a dedicated neural network for each new site.

Where Pith is reading between the lines

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

  • The same RSRP-matching idea could reduce data-collection costs for other site-specific wireless tasks such as beam management or localization.
  • If the calibration memory remains effective across seasonal or construction-induced channel changes, the framework could support longer-term autonomous adaptation in live networks.
  • Combining SiFo-style calibration with larger multimodal foundation models might allow joint optimization of feedback, positioning, and sensing at new sites.

Load-bearing premise

RSRP fingerprints collected during standard SSB probing are distinctive enough to correctly match a served user to nearby calibration samples whose stored CSI labels remain accurate.

What would settle it

In a new deployment, if users matched by RSRP exhibit CSI reconstruction error no better than a non-calibrated model and well below the Type-II reference across multiple locations, the adaptation method would be falsified.

Figures

Figures reproduced from arXiv: 2605.16141 by Cheng-Jie Zhao, Yuanwei Liu, Zhaolin Wang, Zongyao Zhao.

Figure 1
Figure 1. Figure 1: Illustration of the SiFo framework work HΘ⋆ 0 (·) that maps an RSRP fingerprint to a Q￾dimensional CSI representation basis. This stage provides shared RSRP fingerprint coordinates and a transferable parametric prior for CSI subspace acquisition. • Stage 2: Target-site projector calibration. For a new target site t, the BS collects a calibration UE set Ct. Each calibration UE is measured through the pretra… view at source ↗
Figure 3
Figure 3. Figure 3: Comparison of different RSRP sensing coordinates. [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Effective spectral efficiency comparison under limited target [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] view at source ↗
read the original abstract

SiFo, a wireless foundation model-based framework, is proposed for low-overhead site-specific channel state information (CSI) feedback. In 3GPP NR, Type-II feedback provides an expressive codebook-based CSI representation, but it requires substantial reference-signal overhead, UE-side search, and feedback. Learning-based site-specific feedback can reduce these online costs while retaining high-quality subspace representation by exploiting deployment-dependent propagation structure. However, existing site-specific designs typically train a dedicated neural network for each new site, which limits scalability when the number of deployments is large. SiFo addresses this scalability issue by pretraining a CSI feedback model across source sites and adapting it to a target site through lightweight calibration. A small set of target-site users reports low-dimensional reference signal received power (RSRP) fingerprints, and their full-CSI-based subspace labels are stored as calibration memory. During online operation, a served user is matched to calibrated users through the same SSB probing and RSRP reporting procedure, so nearby calibration samples provide site-specific subspace guidance without updating model parameters. SiFo therefore transfers common propagation knowledge while retaining local adaptation. Numerical results across ten city scenarios demonstrate that SiFo (i) achieves higher CSI-capture efficiency than separately trained site-specific learning baselines under the same target-site labeled budget, (ii) approaches the high-overhead 3GPP NR Type-II feedback reference using only RSRP measurements collected during online SSB probing, and (iii) converts the high CSI-capture efficiency and low overhead into effective spectral efficiency improvement under limited target-site data.

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

2 major / 1 minor

Summary. The paper proposes SiFo, a wireless foundation model pretrained across source sites for CSI feedback that adapts to target sites via lightweight calibration: a small set of target-site users report RSRP fingerprints whose associated full-CSI subspace labels are stored in calibration memory. During online operation a served user is matched to the memory via the same SSB RSRP procedure so that the retrieved label supplies site-specific subspace guidance without updating model parameters. Numerical results across ten city scenarios claim that SiFo attains higher CSI-capture efficiency than separately trained site-specific baselines under identical target-site labeled budgets, approaches the high-overhead 3GPP NR Type-II reference using only RSRP measurements, and yields improved spectral efficiency under limited target-site data.

Significance. If the central claims hold, SiFo would provide a scalable route to site-specific CSI feedback by transferring common propagation knowledge while retaining local adaptation through standard SSB probing, avoiding per-deployment retraining. The multi-scenario empirical evaluation and explicit comparison against the 3GPP Type-II reference are concrete strengths that enhance practical relevance.

major comments (2)
  1. [Online operation paragraph] Online-operation paragraph: the headline efficiency and spectral-efficiency gains rest on the assumption that RSRP fingerprints obtained during standard SSB probing are locally discriminative with respect to the dominant eigenspace. No quantitative bound on matching error, no ablation on RSRP dimensionality or beam configuration, and no verification that users sharing similar RSRP vectors share similar subspaces are supplied; if this locality fails, the transferred labels inject irreducible error that the frozen foundation model cannot correct.
  2. [Results section] Results section: the abstract and main claims cite numerical superiority across ten city scenarios, yet the manuscript supplies no details on model architecture, training procedure, exact definition of CSI-capture efficiency, error bars, or data-exclusion rules. Without these elements the reported gains cannot be independently verified and therefore do not yet substantiate the cross-baseline and cross-reference comparisons.
minor comments (1)
  1. [Abstract] Abstract: the term 'CSI-capture efficiency' is introduced without an explicit definition or formula; a concise definition should appear at first use in the main text.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed review of our manuscript. We address each major comment below, indicating where revisions will be made to improve clarity and reproducibility while preserving the core contributions.

read point-by-point responses

  1. Referee: [Online operation paragraph] Online-operation paragraph: the headline efficiency and spectral-efficiency gains rest on the assumption that RSRP fingerprints obtained during standard SSB probing are locally discriminative with respect to the dominant eigenspace. No quantitative bound on matching error, no ablation on RSRP dimensionality or beam configuration, and no verification that users sharing similar RSRP vectors share similar subspaces are supplied; if this locality fails, the transferred labels inject irreducible error that the frozen foundation model cannot correct.

    Authors: We acknowledge that the manuscript does not supply an explicit quantitative bound on RSRP matching error or dedicated ablations on dimensionality and beam configuration. The multi-scenario evaluation provides empirical support for the locality assumption, yet we agree that direct verification would strengthen the claims. In the revised manuscript we will add a dedicated analysis subsection that quantifies the correlation between RSRP-vector similarity and dominant-eigenspace similarity across the ten city scenarios, together with ablations on RSRP dimensionality and SSB beam configurations. This will include verification that users with similar RSRP fingerprints indeed share similar subspaces under the evaluated propagation conditions. revision: yes

  2. Referee: [Results section] Results section: the abstract and main claims cite numerical superiority across ten city scenarios, yet the manuscript supplies no details on model architecture, training procedure, exact definition of CSI-capture efficiency, error bars, or data-exclusion rules. Without these elements the reported gains cannot be independently verified and therefore do not yet substantiate the cross-baseline and cross-reference comparisons.

    Authors: We agree that additional implementation details are required for independent verification. The revised manuscript will expand the relevant sections to provide: the full model architecture and hyper-parameters, the complete training procedure (including loss functions, optimizer settings, and pre-training protocol), the precise mathematical definition of CSI-capture efficiency, error bars or confidence intervals for all plotted metrics, and explicit data-exclusion rules together with scenario-selection criteria. These additions will directly support the cross-baseline and 3GPP Type-II comparisons. revision: yes

Circularity Check

0 steps flagged

Empirical framework evaluated against external 3GPP and per-site baselines; no derivation reduces to fitted inputs or self-citation by construction

full rationale

The paper presents SiFo as a pretraining-plus-calibration framework whose performance claims rest on numerical results across ten city scenarios compared to separately trained site-specific baselines and the 3GPP NR Type-II reference. No equations, uniqueness theorems, or ansatzes are shown that reduce a claimed prediction to a fitted parameter or prior self-citation. The online RSRP-matching procedure is described as an empirical mechanism whose accuracy is assessed via end-to-end spectral-efficiency metrics rather than derived from internal definitions. This is the most common honest finding for an empirical wireless framework whose central results are externally benchmarked.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Based on abstract only; no explicit free parameters, axioms, or invented entities are stated. The framework implicitly assumes shared propagation structure across sites and that RSRP matching suffices for subspace transfer.

pith-pipeline@v0.9.0 · 5818 in / 1265 out tokens · 59542 ms · 2026-05-19T21:42:36.007445+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

33 extracted references · 33 canonical work pages · 1 internal anchor

  1. [1]

    On the Opportunities and Risks of Foundation Models

    R. Bommasaniet al., “On the opportunities and risks of foundation models,”arXiv preprint arXiv:2108.07258, 2021

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  2. [2]

    WirelessLLM: Empowering large language models towards wireless intelligence,

    J. Shao, J. Tong, Q. Wu, W. Guo, Z. Li, Z. Lin, and J. Zhang, “WirelessLLM: Empowering large language models towards wireless intelligence,”J. Commun. Inf. Netw., vol. 9, no. 2, pp. 99–112, Jun. 2024

    work page 2024

  3. [3]

    WiFo: Wireless foundation model for channel prediction,

    B. Liu, S. Gao, X. Liu, X. Cheng, and L. Yang, “WiFo: Wireless foundation model for channel prediction,”Sci. China Inf. Sci., vol. 68, no. 6, p. 162302, May 2025

    work page 2025

  4. [4]

    Large wireless model (LWM): A foundation model for wireless channels,

    S. Alikhani, G. Charan, and A. Alkhateeb, “Large wireless model: Foun- dation model for wireless channels,”arXiv preprint arXiv:2411.08872, 2024

    work page arXiv 2024

  5. [5]

    WirelessGPT: A generative pre-trained multi-task learning framework for wireless communication and sensing,

    T. Yang, P. Zhang, M. Zheng, Y . Shi, L. Jing, J. Huang, and N. Li, “WirelessGPT: A generative pre-trained multi-task learning framework for wireless communication and sensing,”IEEE Netw., vol. 39, no. 5, pp. 58–65, 2025

    work page 2025

  6. [6]

    A wireless foundation model for multi-task prediction,

    Y . Sheng, J. Wang, X. Zhou, L. Liang, H. Ye, S. Jin, and G. Y . Li, “A wireless foundation model for multi-task prediction,”arXiv preprint arXiv:2507.05938, 2025

    work page arXiv 2025

  7. [7]

    6G WavesFM: A multimodal foundation model for sensing, communication, and localiza- tion,

    A. Aboulfotouh, E. Mohammed, and H. Abou-Zeid, “6G WavesFM: A multimodal foundation model for sensing, communication, and localiza- tion,”IEEE Open J. Commun. Soc., vol. 6, pp. 6792–6807, 2025

    work page 2025

  8. [8]

    SpectrumFM: A foundation model for intelligent spectrum management,

    F. Zhou, C. Liu, H. Zhang, W. Wu, Q. Wu, T. Q. S. Quek, and C.- B. Chae, “SpectrumFM: A foundation model for intelligent spectrum management,”IEEE J. Sel. Areas Commun., vol. 44, no. 2, pp. 4471– 4488, 2026

    work page 2026

  9. [9]

    IQFM: A foundation model for wireless multi-antenna IQ streams,

    O. A. Mashaal and H. Abou-Zeid, “IQFM: A foundation model for wireless multi-antenna IQ streams,”IEEE Open J. Commun. Soc., vol. 7, pp. 3483–3501, 2026

    work page 2026

  10. [10]

    WiFo-CF: Wireless foundation model for CSI feedback,

    X. Liu, S. Gao, B. Liu, X. Cheng, and L. Yang, “WiFo-CF: Wireless foundation model for CSI feedback,”IEEE Trans. Wireless Commun., vol. 25, pp. 15 039–15 053, 2026

    work page 2026

  11. [11]

    An overview of limited feedback in wireless commu- nication systems,

    D. J. Love, R. W. Heath, V . K. N. Lau, D. Gesbert, B. D. Rao, and M. Andrews, “An overview of limited feedback in wireless commu- nication systems,”IEEE J. Sel. Areas Commun., vol. 26, no. 8, pp. 1341–1365, Oct. 2008

    work page 2008

  12. [12]

    A tutorial on downlink precoder selection strategies for 3GPP MIMO codebooks,

    X. Fu, D. Le Ruyet, R. Visoz, V . Ramireddy, M. Grossmann, M. Land- mann, and W. Quiroga, “A tutorial on downlink precoder selection strategies for 3GPP MIMO codebooks,”IEEE Access, vol. 11, pp. 138 897–138 922, Dec. 2023

    work page 2023

  13. [13]

    NR; physical layer procedures for data,

    3GPP, “NR; physical layer procedures for data,” 3rd Generation Part- nership Project (3GPP), Technical Specification TS 38.214, 2018

    work page 2018

  14. [14]

    Machine learning codebook design for initial access and CSI type-II feedback in sub-6-GHz 5G NR,

    R. M. Dreifuerst and R. W. Heath, “Machine learning codebook design for initial access and CSI type-II feedback in sub-6-GHz 5G NR,”IEEE Trans. Wireless Commun., vol. 23, no. 6, pp. 6411–6424, Jun. 2024

    work page 2024

  15. [15]

    Neural codebook design for MIMO network beam management,

    ——, “Neural codebook design for MIMO network beam management,” IEEE Trans. Wireless Commun., vol. 24, no. 5, pp. 3909–3922, May 2025

    work page 2025

  16. [16]

    Learning site-specific probing beams for fast mmwave beam alignment,

    Y . Heng, J. Mo, and J. G. Andrews, “Learning site-specific probing beams for fast mmwave beam alignment,”IEEE Trans. Wireless Com- mun., vol. 21, no. 8, pp. 5785–5800, Jan. 2022

    work page 2022

  17. [17]

    Learning beams adaptive to the environment: An RSRP-based code- book design,

    X. Ning, S. Zhang, Y . Xue, X. Zheng, Q. Shi, and T.-H. Chang, “Learning beams adaptive to the environment: An RSRP-based code- book design,” inProc. IEEE Int. Workshop Signal Process. Adv. Wireless Commun. (SPAWC), 2023, pp. 521–525

    work page 2023

  18. [18]

    Environment-aware hybrid beamforming by leveraging channel knowledge map,

    D. Wu, Y . Zeng, S. Jin, and R. Zhang, “Environment-aware hybrid beamforming by leveraging channel knowledge map,”IEEE Trans. Wireless Commun., vol. 23, no. 5, pp. 4990–5005, Oct. 2024

    work page 2024

  19. [19]

    A tutorial on environment-aware communications via channel knowledge map for 6g,

    Y . Zeng, J. Chen, J. Xu, D. Wu, X. Xu, S. Jin, X. Gao, D. Gesbert, S. Cui, and R. Zhang, “A tutorial on environment-aware communications via channel knowledge map for 6g,”IEEE Commun. Surv. Tutorials, vol. 26, no. 3, pp. 1478–1519, Feb 2024

    work page 2024

  20. [20]

    Grid-free MIMO beam alignment through site-specific deep learning,

    Y . Heng and J. G. Andrews, “Grid-free MIMO beam alignment through site-specific deep learning,”IEEE Trans. Wireless Commun., vol. 23, no. 2, pp. 908–921, Feb. 2024

    work page 2024

  21. [21]

    Generative site-specific beamforming for next-generation spatial intelligence,

    Z. Wang, Z. Zhou, C.-J. Zhao, and Y . Liu, “Generative site-specific beamforming for next-generation spatial intelligence,”arXiv preprint arXiv:2601.02301, 2026

    work page arXiv 2026

  22. [22]

    Generative site-specific beam- forming via information-maximizing codebook,

    C.-J. Zhao, Z. Wang, and Y . Liu, “Generative site-specific beam- forming via information-maximizing codebook,”arXiv preprint arXiv:2602.12552, 2026

    work page arXiv 2026

  23. [23]

    Bridging standardized codebook and site-specific beamforming: A unified limited-feedback framework,

    C.-J. Zhao, Z. Wang, Z. Zhao, and Y . Liu, “Bridging standardized codebook and site-specific beamforming: A unified limited-feedback framework,” 2026, manuscript

    work page 2026

  24. [24]

    Spatially sparse precoding in millimeter wave MIMO systems,

    O. E. Ayach, S. Rajagopal, S. Abu-Surra, Z. Pi, and R. W. Heath, “Spatially sparse precoding in millimeter wave MIMO systems,”IEEE Trans. Wireless Commun., vol. 13, no. 3, pp. 1499–1513, Mar. 2014

    work page 2014

  25. [25]

    Aspects of favorable propagation in massive MIMO,

    H. Q. Ngo, E. G. Larsson, and T. L. Marzetta, “Aspects of favorable propagation in massive MIMO,” inProc. 22nd Eur. Signal Process. Conf. (EUSIPCO), 2014, pp. 76–80

    work page 2014

  26. [26]

    Massive MIMO for next generation wireless systems,

    E. G. Larsson, O. Edfors, F. Tufvesson, and T. L. Marzetta, “Massive MIMO for next generation wireless systems,”IEEE Commun. Mag., vol. 52, no. 2, pp. 186–195, Feb. 2014

    work page 2014

  27. [27]

    NR; physical channels and modulation,

    3GPP, “NR; physical channels and modulation,” 3rd Generation Part- nership Project (3GPP), Technical Specification TS 38.211, 2018

    work page 2018

  28. [28]

    NR; physical layer measurements,

    ——, “NR; physical layer measurements,” 3rd Generation Partnership Project (3GPP), Technical Specification TS 38.215, 2018

    work page 2018

  29. [29]

    A tutorial on beam management for 3GPP NR at mmwave frequencies,

    M. Giordani, M. Polese, A. Roy, D. Castor, and M. Zorzi, “A tutorial on beam management for 3GPP NR at mmwave frequencies,”IEEE Commun. Surv. Tutorials, vol. 21, no. 1, pp. 173–196, Sep. 2019

    work page 2019

  30. [30]

    DeepMIMO: A generic deep learning dataset for mil- limeter wave and massive MIMO applications,

    A. Alkhateeb, “DeepMIMO: A generic deep learning dataset for mil- limeter wave and massive MIMO applications,” inProc. Inf. Theory Appl. Workshop (ITA), San Diego, CA, Feb. 2019, pp. 1–8

    work page 2019

  31. [31]

    Adam: A method for stochastic optimization,

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” inProc. Int. Conf. Learn. Represent. (ICLR), 2015

    work page 2015

  32. [32]

    Explicit inductive bias for transfer learning with convolutional networks,

    X. Li, Y . Grandvalet, and F. Davoine, “Explicit inductive bias for transfer learning with convolutional networks,” inProc. Int. Conf. Mach. Learn. (ICML), 2018, pp. 2825–2834

    work page 2018

  33. [33]

    R. A. Horn and C. R. Johnson,Matrix Analysis, 2nd ed. Cambridge University Press, 2012

    work page 2012