arxiv: 2605.03279 · v1 · submitted 2026-05-05 · 💻 cs.LG

Recognition: unknown

RFPrompt: Prompt-Based Expert Adaptation of the Large Wireless Model for Modulation Classification

Md Raihan Uddin , Tolunay Seyfi , Fatemeh Afghah

Authors on Pith no claims yet

Pith reviewed 2026-05-07 17:40 UTC · model grok-4.3

classification 💻 cs.LG

keywords prompt-based adaptationautomatic modulation classificationwireless foundation modelsdistribution shiftparameter-efficient adaptationover-the-air IQ datalarge wireless model

0 comments

The pith

Prompt tokens steer a frozen wireless foundation model to classify modulations more robustly under real-world shifts.

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

The paper tests whether small learnable prompt tokens can adapt a pretrained large wireless model to automatic modulation classification without retraining the entire network. It focuses on cases where test signals come from different hardware, channels, or recording setups than the original training data. Only the prompt tokens are updated while the backbone stays fixed, which keeps the number of new parameters very low. This matters for wireless systems that must operate in changing environments where full retraining is impractical and labeled data is scarce. The results indicate the prompt method improves accuracy on both synthetic shifts and actual over-the-air recordings.

Core claim

RFPrompt inserts learnable deep prompt tokens into the frozen Large Wireless Model so that only those tokens are trained for the modulation classification task; this yields better robustness to distribution shifts and limited supervision on real-world IQ data while preserving parameter efficiency.

What carries the argument

Learnable deep prompt tokens inserted into the frozen Large Wireless Model backbone to steer its existing representations toward a specific downstream task.

Load-bearing premise

The representations already learned by the Large Wireless Model during pretraining remain useful and can be redirected by added prompt tokens without losing their general value.

What would settle it

A test showing that the prompt-adapted model produces lower classification accuracy than either the frozen base model or a fully fine-tuned version on held-out real over-the-air IQ recordings would falsify the claimed benefit.

Figures

Figures reproduced from arXiv: 2605.03279 by Fatemeh Afghah, Md Raihan Uddin, Tolunay Seyfi.

**Figure 1.** Figure 1: End-to-end overview of the proposed RFPrompt system. view at source ↗

**Figure 2.** Figure 2: t-SNE projections of router-fused embeddings view at source ↗

read the original abstract

Automatic modulation classification (AMC) in real-world deployments demands robustness to distribution shifts arising from hardware impairments, unseen propagation environments, and recording conditions never encountered during training. Although wireless foundation models offer a promising starting point for robust RF representation learning, an important open question is how to adapt them efficiently to out-of-distribution (OOD) downstream tasks without overwriting the structure learned during large-scale pre-training. In this paper, we investigate prompt-based adaptation as a general mechanism for OOD transfer in wireless foundation models. We propose RFPrompt, a parameter-efficient framework that introduces learnable deep prompt tokens while keeping the pretrained backbone frozen, enabling task-specific adaptation with minimal trainable parameters. We instantiate and evaluate this approach on the Large Wireless Model (LWM), a mixture-of-experts wireless foundation model, and study its behavior under both standard and OOD modulation-classification settings. Results show that prompt-based adaptation consistently improves robustness under distribution shift and limited supervision, particularly on real-world over-the-air IQ data, while preserving strong parameter efficiency. These findings suggest that prompt learning is a practical and effective strategy for adapting wireless foundation models to challenging downstream RF environments.

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.

Desk Editor's Note private letter to a colleague

RFPrompt shows that adding learnable deep prompts to a frozen wireless MoE model improves OOD modulation classification on real OTA data with high parameter efficiency, but the experiments do not isolate whether the prompts steer the backbone representations or simply supply an extra input channel.

read the letter

The main takeaway is that this paper takes prompt tuning, freezes the Large Wireless Model backbone, and adds trainable deep prompt tokens to handle distribution shifts in automatic modulation classification. It reports better robustness under limited supervision and real over-the-air conditions while keeping trainable parameters very low. That setup is the core contribution and the part worth paying attention to if you work in wireless ML adaptation.

Referee Report

3 major / 2 minor

Summary. The paper proposes RFPrompt, a parameter-efficient prompt-based adaptation framework for the Large Wireless Model (LWM), a mixture-of-experts wireless foundation model, applied to automatic modulation classification (AMC). Learnable deep prompt tokens are introduced while the pretrained backbone remains frozen, with the goal of enabling robust adaptation to out-of-distribution tasks arising from hardware impairments, unseen propagation environments, and real-world over-the-air IQ data under limited supervision.

Significance. If the central empirical claims hold after addressing the controls below, the work would demonstrate a practical route for adapting large wireless foundation models to distribution-shifted RF tasks without full retraining, preserving parameter efficiency and pre-trained structure. This is relevant given the scale of such models and the prevalence of OOD conditions in real deployments; the emphasis on real OTA data and mixture-of-experts backbones is a concrete strength.

major comments (3)

[Results] Results section: The manuscript does not report feature-space similarity metrics or expert-activation statistics comparing the frozen backbone before and after prompt adaptation. These diagnostics are required to substantiate the claim that learnable prompt tokens steer general RF representations rather than simply acting as an auxiliary input channel that the backbone exploits differently.
[Experiments] Experiments section: No post-training ablation is presented in which the learned prompt tokens are removed or frozen while keeping the backbone fixed; such an ablation is load-bearing for isolating whether performance gains on OOD real-world OTA IQ data arise specifically from the prompt-based adaptation mechanism.
[Experiments] Experiments section: The evaluation omits a matched-parameter unfrozen fine-tuning baseline (i.e., updating a comparable number of parameters in the backbone). Without this comparison, it is impossible to determine whether the reported robustness improvements under distribution shift are attributable to the frozen-plus-prompt design or could be achieved by standard adaptation with equivalent compute.

minor comments (2)

[Abstract] Abstract: The statement that 'results show that prompt-based adaptation consistently improves robustness' supplies no numerical values, baselines, dataset sizes, or error bars, which reduces the abstract's utility for quick assessment of the claims.
Ensure that the specific OOD scenarios, real-world OTA dataset characteristics, and exact prompt insertion points within the MoE architecture are described with sufficient detail to support reproducibility.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. The suggestions for additional diagnostics and baselines will help strengthen the validation of the prompt adaptation mechanism. We address each major comment below and will incorporate the requested analyses into the revised manuscript.

read point-by-point responses

Referee: [Results] Results section: The manuscript does not report feature-space similarity metrics or expert-activation statistics comparing the frozen backbone before and after prompt adaptation. These diagnostics are required to substantiate the claim that learnable prompt tokens steer general RF representations rather than simply acting as an auxiliary input channel that the backbone exploits differently.

Authors: We agree that these diagnostics would provide stronger evidence that the prompts actively steer the representations. In the revised manuscript, we will add expert-activation statistics from the mixture-of-experts layers, comparing activation patterns on OOD samples with and without the learned prompts. We will also report feature-space similarity metrics (e.g., cosine similarity of embeddings) to quantify changes induced by the prompts. revision: yes
Referee: [Experiments] Experiments section: No post-training ablation is presented in which the learned prompt tokens are removed or frozen while keeping the backbone fixed; such an ablation is load-bearing for isolating whether performance gains on OOD real-world OTA IQ data arise specifically from the prompt-based adaptation mechanism.

Authors: This is a valid concern for isolating the mechanism. We will include the requested post-training ablation in the revision: after training, we will evaluate performance when replacing the learned prompt tokens with random initialization (while keeping the backbone frozen) and compare against the trained prompts. This will confirm that gains on real-world OTA data derive from the optimized prompts rather than the auxiliary input structure alone. revision: yes
Referee: [Experiments] Experiments section: The evaluation omits a matched-parameter unfrozen fine-tuning baseline (i.e., updating a comparable number of parameters in the backbone). Without this comparison, it is impossible to determine whether the reported robustness improvements under distribution shift are attributable to the frozen-plus-prompt design or could be achieved by standard adaptation with equivalent compute.

Authors: We acknowledge that this baseline is necessary to contextualize the benefits. In the revised manuscript, we will add results from a matched-parameter unfrozen fine-tuning baseline, where we update a comparable number of backbone parameters (e.g., via selective layer unfreezing or low-rank updates matching the prompt token count) and evaluate OOD robustness on the same tasks. This will clarify whether the frozen-plus-prompt design offers advantages beyond equivalent compute. revision: yes

Circularity Check

0 steps flagged

Empirical evaluation of prompt adaptation with no circular derivations

full rationale

The paper is an empirical study proposing and evaluating RFPrompt for adapting a pretrained Large Wireless Model (LWM) to modulation classification under distribution shifts, using a frozen backbone plus learnable prompts. No mathematical derivations, equations, or predictions appear in the abstract or described content. Claims rest on experimental results on standard and OOD tasks including real OTA IQ data, with no self-referential reductions, fitted inputs renamed as predictions, or load-bearing self-citations that collapse the central findings to prior inputs by construction. The work is self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

Review based on abstract only; full details on parameters, assumptions, and methods unavailable. The central claim rests on the effectiveness of prompt tokens for adaptation.

free parameters (1)

learnable deep prompt tokens
These are the trainable parameters introduced for task-specific adaptation; their count, initialization, and placement are not specified in the abstract.

axioms (1)

domain assumption The Large Wireless Model has learned general RF representations that remain useful when the backbone is frozen during prompt adaptation
This underpins the decision to keep the pretrained model fixed while only training prompts.

pith-pipeline@v0.9.0 · 5507 in / 1440 out tokens · 105455 ms · 2026-05-07T17:40:49.568930+00:00 · methodology

Review history (2 revisions) →

discussion (0)

Reference graph

Works this paper leans on

27 extracted references · 12 canonical work pages · 2 internal anchors

[1]

Over-the-air deep learning based radio signal classification,

T. J. OShea, T. Roy, and T. C. Clancy, “Over-the-air deep learning based radio signal classification,”IEEE Journal of Selected Topics in Signal Processing, vol. 12, no. 1, pp. 168–179, 2018

2018
[2]

Adapt under attack and domain shift: Unified adversarial meta-learning and domain adap- tation for robust automatic modulation classification,

A. Owfi, A. Bamdad, T. Seyfi, and F. Afghah, “Adapt under attack and domain shift: Unified adversarial meta-learning and domain adap- tation for robust automatic modulation classification,”arXiv preprint arXiv:2511.01172, 2025

work page arXiv 2025
[3]

Iqfm–a wireless foundation model for i/q streams in ai-native 6g,

O. Mashaal and H. Abou-Zeid, “Iqfm–a wireless foundation model for i/q streams in ai-native 6g,”IEEE Open Journal of the Communications Society, 2026

2026
[4]

Multimodal wireless foundation models,

A. Aboulfotouh and H. Abou-Zeid, “Multimodal wireless foundation models,”arXiv preprint arXiv:2511.15162, 2025

work page arXiv 2025
[5]

Spectrumfm: A foundation model for intelligent spectrum management,

F. Zhou, C. Liu, H. Zhang, W. Wu, Q. Wu, T. Q. Quek, and C.-B. Chae, “Spectrumfm: A foundation model for intelligent spectrum management,” IEEE Journal on Selected Areas in Communications, 2025

2025
[6]

RF-GPT: Teaching AI to see the wireless world,

H. Zou, Y . Tian, B. Wang, L. Bariah, S. Lasaulce, C. Huang, and M. Debbah, “Rf-gpt: Teaching ai to see the wireless world,”arXiv preprint arXiv:2602.14833, 2026, version v1

work page arXiv 2026
[7]

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

S. Alikhani, G. Charan, and A. Alkhateeb, “Large wireless model (lwm): A foundation model for wireless channels,”arXiv preprint arXiv:2411.08872, 2024

work page arXiv 2024
[8]

An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale

A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gellyet al., “An image is worth 16x16 words: Transformers for image recognition at scale,”arXiv preprint arXiv:2010.11929, 2020

work page internal anchor Pith review arXiv 2010
[9]

Lwm-spectro: A foundation model for wireless baseband signal spectrograms,

N. Kim, S. Alikhani, and A. Alkhateeb, “Lwm-spectro: A foundation model for wireless baseband signal spectrograms,”arXiv preprint arXiv:2601.08780, 2026, version v1

work page arXiv 2026
[10]

Lwm- temporal: Sparse spatio-temporal attention for wireless channel represen- tation learning,

S. Alikhani, A. Malhotra, S. Hamidi-Rad, and A. Alkhateeb, “Lwm- temporal: Sparse spatio-temporal attention for wireless channel represen- tation learning,”arXiv preprint arXiv:2603.10024, 2026

work page arXiv 2026
[11]

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,”arXiv preprint arXiv:1902.06435, 2019

work page Pith review arXiv 1902
[12]

Adaptive meta-learning-based adversarial training for robust automatic modulation classification,

A. Bamdad, A. Owfi, and F. Afghah, “Adaptive meta-learning-based adversarial training for robust automatic modulation classification,” in 2025 IEEE International Conference on Communications Workshops (ICC Workshops). IEEE, 2025, pp. 292–297

2025
[13]

Metaamc: meta learning and automl for model compression,

C. Zhang, Y . Zhu, and Z. Bai, “Metaamc: meta learning and automl for model compression,” inTwelfth International Conference on Digital Image Processing (ICDIP 2020), vol. 11519. SPIE, 2020, pp. 536–541

2020
[14]

Visual prompt tuning,

M. Jia, L. Tang, B.-C. Chen, C. Cardie, S. Belongie, B. Hariharan, and S.-N. Lim, “Visual prompt tuning,” inEuropean conference on computer vision. Springer, 2022, pp. 709–727

2022
[15]

Learning to prompt for continual learning,

Z. Wang, Z. Zhang, C.-Y . Lee, H. Zhang, R. Sun, X. Ren, G. Su, V . Perot, J. Dy, and T. Pfister, “Learning to prompt for continual learning,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 139–149

2022
[16]

Lora: Low-rank adaptation of large language models

E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, W. Chenet al., “Lora: Low-rank adaptation of large language models.” Iclr, vol. 1, no. 2, p. 3, 2022

2022
[17]

pmoe: Prompting diverse experts together wins more in visual adaptation,

S. Mo, X. Luo, and D. Li, “pmoe: Prompting diverse experts together wins more in visual adaptation,”arXiv preprint arXiv:2602.22938, 2026

work page arXiv 2026
[18]

Revisit visual prompt tuning: The expressiveness of prompt experts,

M. Le, A. Nguyen, H. Nguyen, C. Nguyen, A. T. Tran, and N. Ho, “Revisit visual prompt tuning: The expressiveness of prompt experts,” inThe Fourteenth International Conference on Learning Representations, 2026. [Online]. Available: https://openreview.net/forum? id=EbjbESm8MD

2026
[19]

Dataset: Iq samples of lte, 5g nr, wi-fi, its-g5, and c-v2x pc5,

M. Girmay and A. Shahid, “Dataset: Iq samples of lte, 5g nr, wi-fi, its-g5, and c-v2x pc5,” 2023. [Online]. Available: https://dx.doi.org/10.21227/72qq-z464

work page doi:10.21227/72qq-z464 2023
[20]

Real-world iq dataset for automatic radio modulation recognition under multipath channels,

N. Belousov and M. Ronkin, “Real-world iq dataset for automatic radio modulation recognition under multipath channels,” 2026. [Online]. Available: https://data.mendeley.com/datasets/tjzsbph49x/2

2026
[21]

Deep residual learning for image recognition,

K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 770–778

2016
[22]

Efficientnet: Rethinking model scaling for convolu- tional neural networks,

M. Tan and Q. Le, “Efficientnet: Rethinking model scaling for convolu- tional neural networks,” inInternational conference on machine learning. PMLR, 2019, pp. 6105–6114

2019
[23]

Searching for mobilenetv3,

A. Howard, M. Sandler, G. Chu, L.-C. Chen, B. Chen, M. Tan, W. Wang, Y . Zhu, R. Pang, V . Vasudevanet al., “Searching for mobilenetv3,” in Proceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 1314–1324

2019
[24]

Radio modulation classification using stft spectrogram and cnn,

J. Wu, Y . Zhong, and A. Chen, “Radio modulation classification using stft spectrogram and cnn,” in2021 7th International Conference on Computer and Communications (ICCC). IEEE, 2021, pp. 178–182

2021
[25]

Model-agnostic meta-learning for fast adaptation of deep networks,

C. Finn, P. Abbeel, and S. Levine, “Model-agnostic meta-learning for fast adaptation of deep networks,” inInternational conference on machine learning. PMLR, 2017, pp. 1126–1135

2017
[26]

On First-Order Meta-Learning Algorithms

A. Nichol and J. Schulman, “Reptile: a scalable metalearning algorithm,” arXiv preprint arXiv:1803.02999, vol. 2, no. 3, p. 4, 2018

work page Pith review arXiv 2018
[27]

Decoupled Weight Decay Regularization

I. Loshchilov and F. Hutter, “Decoupled weight decay regularization,” arXiv preprint arXiv:1711.05101, 2017

work page internal anchor Pith review arXiv 2017