PaperlayerView exact revision on arXiv

Source-linked AI summary

RadioLLM: Introducing Large Language Model into Cognitive Radio via Hybrid Prompt and Token Reprogrammings

Shuai Chen, Yong Zu, Zhixi Feng, Shuyuan Yang, Mengchang Li

arXiv:2501.17888v3eess.SPcs.AIcs.LG

TL;DR

Existing CRT frameworks are often task-specific, and fixed prompts can consume tokens and dilute signal features. RadioLLM combines hybrid prompt and token reprogramming with high-frequency feature fusion; experiments across diverse tasks validate its effectiveness, and the joint HTRP–FAF configuration reports OA of 58.10%.

  • Problem

    DL-based CRT architectures can be task- and signal-type-specific, while fixed prompts consume prefix tokens and dilute signal features.

    Source: §I · p. 1

  • Method

    RadioLLM combines hybrid prompt and token reprogramming to join radio signal features with expert knowledge, and uses FAF to model high-frequency information.

    Source: §I · p. 2

  • Results

    Experiments across diverse tasks validate RadioLLM’s effectiveness; the joint HTRP–FAF configuration achieves OA of 58.10%.

    Source: §I · p. 11, §V · p. 12

  • Takeaways & Limitations

    RadioLLM is presented as a framework for versatile CRT applications across diverse scenarios through multi-task joint learning.

    Source: §I · p. 2

Abstract

from arXiv · show

The growing scarcity of spectrum resources and rapid proliferation of wireless devices make efficient radio network management critical. While deep learning-enhanced Cognitive Radio Technology (CRT) provides promising solutions for tasks such as radio signal classification (RSC), denoising, and spectrum allocation, existing DL-based CRT frameworks are typically task-specific and lack scalability in diverse real-world applications. This limitation naturally leads to the exploration of Large Language Models (LLMs), whose exceptional cross-domain generalization capabilities offer new potential for advancing CRT. To bridge this gap, we propose RadioLLM, a novel framework that integrates Hybrid Prompt and Token Reprogramming (HPTR) for combining radio signal features with expert knowledge, and a Frequency-Attuned Fusion (FAF) module for enhanced high-frequency feature modeling. Extensive evaluations on multiple benchmark datasets demonstrate that RadioLLM achieves superior performance compared to existing baselines in the majority of testing scenarios.

I. INTRODUCTION

RadioLLM addresses the limited scalability of task-specific CRT frameworks amid growing spectrum scarcity and wireless-device proliferation. It combines signal-oriented LLM reprogramming and hybrid prompts with high-frequency feature fusion, against a backdrop of DL approaches that remain tied to particular tasks and signal types.

I. INTRODUCTION

  • DL-based CRT architectures remain task- and signal-type-specific, constraining deployment in heterogeneous environments and motivating a more universal framework.Traditional ML-based radio cognition is also hindered by device-intrinsic noise and increasing interference sources.Source: §I · p. 1
  • LLMs offer cross-domain generalization potential for CRT, but their deployment requires addressing substantial computational demands and adapting architectures to radio signals.The paper frames LLMs as promising CRT candidates while noting these deployment challenges.Source: §I · p. 1
  • Text descriptions can discard signal features and block direct spatiotemporal learning, while fixed expert prompts consume prefix tokens and dilute signal representations.RadioLLM instead maps raw I/Q sequences into LLM-compatible embeddings and uses semantic anchors in place of hardware-fixed prompts.Source: §I · pp. 1–2
  • RadioLLM integrates hybrid prompt and token reprogramming to couple radio features with expert knowledge, and adds FAF to model high-frequency information.FAF combines CNN-extracted local features with LLM global context to improve RSC performance.Source: §I · p. 2

A. Traditional DL-Based CRT Framework

  • Prior work applies CNN/GRU and multi-view fusion to RSC,, while self-supervised RSC and denoising methods address specific tasks,,,.These approaches excel in individual applications but remain limited to particular tasks and data types.Source: A. Traditional DL-Based CRT Framework · p. 2

B. Previous LLM-Based CRT Framework

Earlier LLM applications in telecommunications largely focus on language understanding, while existing CRT systems have limitations in task coverage and unified representation. RadioLLM combines hybrid prompting and signal reprogramming in a framework that maps received signals to task-specific targets such as reconstructed signals or class labels.

B. Previous LLM-Based CRT Framework

  • WirelessLLM depends on external tools and omits WSC and denoising, while LLM4WM’s task-specific adapters impede unified representation and model coherence.Other telecommunications LLM applications are described as mostly limited to language understanding.Source: B. Previous LLM-Based CRT Framework · p. 2
  • RadioLLM uses model reprogramming and hybrid prompting to build a unified CRT framework for diverse signals and tasks.Its overview combines expert prompts with top-K semantically similar anchors to form prefix prompts for signal embeddings, then fuses high- and low-frequency information before the LLM.Source: B. Previous LLM-Based CRT Framework · p. 2, §III · pp. 2–3, Figure 2 · p. 3

A. Problem Statement

  • The received IQ signal is modeled as distorted by the wireless channel and environmental noise before the framework maps it to a task target.The passage identifies channel response and additive white Gaussian noise as components of the received-signal model.Source: A. Problem Statement · p. 3
  • The framework’s target T is a reconstructed signal or task-specific attribute for reconstruction tasks, and a class label for classification.The model is described as mapping received signal r(t) to T using parameters Θ and a task-specific loss.Source: A. Problem Statement · p. 3

B. Hybrid Prompt and Token Reprogramming

RadioLLM combines hardware prompts with a reduced set of semantically selected software prompts, then reprograms signal features into tokens the LLM can process. This design pairs domain-relevant context with radio-signal representations.

B. Hybrid Prompt and Token Reprogramming

  • The hybrid prompt combines verbose hardware prompts with a reduced set of pretrained software prompts to target efficient, domain-relevant representations.The software prompts are selected as semantic anchors from pretrained word-token embeddings.Source: B. Hybrid Prompt and Token Reprogramming · p. 3
  • For each hardware prompt, the method selects the top-K semantic anchors from the reduced embedding set using maximum cosine similarity over prompt tokens.The selected embeddings form the hybrid prompt.Source: B. Hybrid Prompt and Token Reprogramming · p. 3
  • Multi-head cross-attention uses signal features as queries and semantic anchors as keys and values, producing signal tokens projected to the LLM-compatible dimension.The resulting tokens are fed into the LLM for further processing.Source: B. Hybrid Prompt and Token Reprogramming · p. 4

C. Frequency Attuned Fusion

Frequency-Attuned Fusion (FAF) augments signal tokens with CNN-extracted high-frequency features alongside global low-frequency information. The output projection decodes LLM features, while training largely freezes the LLM and adapts downstream tasks with task-specific heads or outputs.

C. Frequency Attuned Fusion

  • FAF combines CNN-extracted high-frequency features with signal information to augment tokens, addressing LLM attention’s lower sensitivity to high-frequency details.Three high-frequency extraction layers use convolution, ReLU, and pooling; the enhanced tokens combine global low-frequency information with fine-grained details.Source: C. Frequency Attuned Fusion · p. 4

D. Output Projection

  • The output projection passes LLM features to a decoder, which maps them to signal outputs using either a linear layer or a shallow Transformer decoder.Pretraining minimizes mean squared error between the decoder output and the target signal.Source: D. Output Projection · p. 4

E. Training Strategy

  • Pretraining updates only a subset of mostly frozen LLM parameters via LoRA, while a dataset loss-balancing factor normalizes gradient contributions.For classification, pooled LLM features feed a linear head; for denoising or reconstruction, the decoder output is the prediction.Source: E. Training Strategy · p. 4

IV. EXPERIMENTS

The experiments use seven publicly available datasets, with a defined train-validation-test split and a mixed-data pretraining strategy. Prompts combine dataset and task descriptions with statistics generated for each input instance.

A. Data Overview

  • Seven public datasets are split into training, validation, and test sets in an 8:1:1 ratio for evaluation.The datasets are RML16A, RML16B, RML16C, RML22, RML18A, ADS-B, and Wi-Fi.Source: A. Data Overview · p. 5
  • Pretraining uses high-SNR RML-series samples, the full ADS-B training set, and a random 5% of Wi-Fi training samples.RML samples are selected at SNR ≥14 dB; ADS-B is unfiltered to capture real-world variations, while the Wi-Fi subset is used for computational efficiency.Source: A. Data Overview · p. 5
  • Each prompt combines a dataset description, task description, and input statistics, with the statistics dynamically generated for each instance.Descriptors include signal-domain attributes and statistics such as minimum, maximum, median, trend, and top lag features.Source: A. Data Overview · p. 5
  • The unified, extensible prompt design supports heterogeneous datasets and diverse tasks using both global and instance-level context.Fig. 3 illustrates the color-coded prompt components: domain-specific descriptions, statistical attributes, and task-specific instructions.Source: A. Data Overview · p. 5, Figure 3 · p. 5

B. Dataset Details

The dataset overview spans simulated modulation benchmarks, real-world signal collections, and over-the-air Wi-Fi transmissions. The section also specifies implementation settings and two signal augmentations: phase rotation and time reversal.

B. Dataset Details

  • RML16A, RML16B, and RML16C contain 220,000, approximately 1.2 million, and 162,060 signals, respectively, with 128 I/Q points per sample.RML16B omits AM-SSB relative to RML16A, while RML16C uses the same modulation types as RML16A.Source: B. Dataset Details · p. 5
  • RML22 spans −20 dB to 20 dB SNR, while the Wi-Fi dataset consists of over-the-air transmissions captured using 16 USRP X310 devices at a 2-foot transmitter–receiver distance.RML22 has 420,000 samples; Wi-Fi signals are processed into sequences of 128 complex-valued I/Q samples.Source: B. Dataset Details · p. 5
  • RML18A is the largest listed modulation benchmark, with 2,555,904 samples across 24 modulation types and 26 SNR levels; ADS-B contains 63,105 real-world signals across 198 classes.RML18A samples have 1024 I/Q points; ADS-B signals are sampled at 50 MHz and contain 3000 points.Source: B. Dataset Details · p. 5

C. Implementation Details

  • Pretraining uses AdamW with an initial learning rate of 5 × 10−5, a 50-epoch cap, and early stopping after 20 consecutive epochs without validation-loss improvement.Downstream tasks use cosine annealing; inference time is measured in seconds per batch at batch size 128.Source: C. Implementation Details · p. 6

1) Phase Rotation:

  • Phase rotation multiplies the original signal by a complex phase factor, with the rotation angle typically chosen randomly from [0, 2π).The result is the phase-rotated signal.Source: 1) Phase Rotation · p. 6

2) Signal Reverse:

  • Signal reversal flips the signal along the time axis; its discrete form uses the signal length and discrete time index.The text identifies the original and reversed signals as the operation’s input and output.Source: 2) Signal Reverse · p. 6

3) Time Warp:

Time Warp augments signals by changing temporal dynamics through nonlinear time-axis transformations; evaluation metrics cover classification and reconstruction quality.

3) Time Warp:

  • Time warping enhances signals by altering temporal dynamics through nonlinear transformations of the time axis.For discrete signals, the operation uses a discrete time-warping function.Source: 3) Time Warp · p. 6
  • The warping function maps times within the signal’s total duration, [0, T].Here, T is the signal’s total duration.Source: 3) Time Warp · p. 6

E. Evaluation Metrics

  • OA and Kappa assess classification, while SSIM assesses structural similarity between predicted and ground-truth signals.OA measures the proportion of correctly classified samples.Source: E. Evaluation Metrics · p. 6

2) Kappa Score:

RadioLLM’s classification results are strongest across most evaluated datasets, with performance varying by dataset, sample availability, and SNR. Kappa measures agreement beyond chance, while SSIM evaluates structural similarity for denoising or reconstruction.

2) Kappa Score:

  • Kappa measures classification reliability while accounting for chance agreement; its range is −1 to 1, with 1 indicating perfect agreement.Zero represents chance-level agreement, while negative values indicate systematic disagreement.Source: 2) Kappa Score · p. 6

3) SSIM:

  • SSIM is a perceptual measure of similarity between signals, used to assess denoised or reconstructed results against their originals.SSIM approaches 1 for identical signals.Source: 3) SSIM · p. 6

F. Comparison with RSC methods

  • RadioLLM records the highest OA on six datasets: 58.10% (RML16A), 58.35% (RML16B), 68.19% (RML16C), 59.39% (RML22), 52.03% (RML18A), and 35.41% (Wi-Fi).Table I compares classification methods; bold marks the best performance and underlining the second-best.Source: F. Comparison with RSC methods · p. 7, Table I · p. 7
  • Confusions remain between similar modulation types, including 16QAM and 64-QAM on RML16A and RML16B, and higher-order QAM classes on RML18A.On RML16A, 26.4% of QPSK samples are misclassified as AM-SSB.Source: F. Comparison with RSC methods · pp. 7, 9
  • Without RML22 pretraining, RadioLLM surpasses TcssAMR by 1.22% in OA and achieves the highest Kappa; it also performs strongly on RML18A and Wi-Fi.On ADS-B, RadioLLM does not achieve the best performance, though its results are described as acceptable.Source: F. Comparison with RSC methods · p. 8
  • RadioLLM’s few-shot advantage is most evident with very few samples on RML16A and RML16B, and grows as sample sizes increase.On RML16B, its low-sample advantage over TcssAMR and SemiAMC is slight.Source: F. Comparison with RSC methods · pp. 7–8
  • RadioLLM performs strongly across SNRs, with especially large high-SNR gains on RML18A and RML16B and low-SNR robustness on RML16C and RML22.The reported high-SNR gains are around 10% over the next-best method on RML18A above 12 dB, and approximately 3% on RML16B.Source: F. Comparison with RSC methods · pp. 7, 9
  • The authors attribute RadioLLM’s results to large-scale pretraining and fine-tuning, together with layered feature extraction and attention-based modeling.The evaluation summary reports accuracy, generalization across datasets, and resilience under challenging noise, while noting remaining confusion among similar classes.Source: F. Comparison with RSC methods · p. 9

G. Performance Analysis of Baselines with Multi-Domain Pretraining

The section evaluates multi-domain pretraining of TcssAMR and SemiAMC, finding improvements in many scenarios alongside occasional drops. It also compares denoising outputs on high-SNR samples, where RadioLLM better preserves signal structure visually.

G. Performance Analysis of Baselines with Multi-Domain Pretraining

  • TcssAMR and SemiAMC were pretrained on aggregated RML16A, RML16B, and RML16C data, then fine-tuned separately on each dataset.The experiment probes whether exposure to diverse domains improves generalization and downstream performance.Source: G. Performance Analysis of Baselines… · p. 9
  • Multi-domain pretraining generally improves OA and Kappa for both baselines on some datasets, but causes slight drops on others.The authors attribute some less optimal adaptation to domain-specific biases that may not align with the target dataset.Source: G. Performance Analysis of Baselines… · p. 9
  • Broader pretraining improves adaptability and transfer across scenarios, but optimal real-world signal processing may still require architectural innovation or domain-specific fine-tuning.The authors identify domain mismatch as a risk to mitigate.Source: G. Performance Analysis of Baselines… · pp. 9–10
  • On randomly selected high-SNR samples from three RML datasets, RadioLLM’s denoised waveforms are visually closer to the originals and preserve I/Q structure better.The comparison includes SGFilter and DNCNet alongside RadioLLM.Source: Figure 6 · p. 10

H. Comparison with Denoise methods

RadioLLM’s denoising results are evaluated under added test noise and show stronger structural similarity than the compared baselines across three datasets. Qualitative comparisons also report closer waveform recovery with less distortion.

H. Comparison with Denoise methods

  • RadioLLM’s denoised waveforms are visually closer to the originals, while SGFilter oversmooths and DNCNet leaves distortion or residual artifacts.The qualitative comparison uses randomly selected high-SNR samples from each dataset.Source: H. Comparison with Denoise methods · p. 10
  • RadioLLM achieves SSIM values of 0.838, 0.893, and 0.846 on RML16A, RML16B, and RML16C, respectively, leading all evaluated methods.The comparison tests models under SNRs from 0 dB to 10 dB.Source: Table II · p. 10, H. Comparison with Denoise methods · p. 10
  • SGFilter has the second-best SSIM scores but its fixed, non-adaptive behavior limits handling of complex noise patterns.It performs reasonably on simpler noise profiles.Source: H. Comparison with Denoise methods · p. 10
  • DNCNet scores below both SGFilter and RadioLLM on all three datasets, and the passage attributes this to difficulty capturing multi-scale temporal and frequency dependencies.Source: H. Comparison with Denoise methods · p. 10
  • The authors attribute RadioLLM’s denoising performance to hierarchical relational modeling and frequency-aware fusion that extract relevant signal features.The text describes the findings as supported by both qualitative and quantitative analyses.Source: H. Comparison with Denoise methods · p. 10

I. Ablation Studies

The ablations assess HTRP and FAF individually and jointly, while parameter sensitivity experiments examine Top-K, decoder choice, and LLM choice. The reported settings include RML16A and a consistent 100-shot setup for sensitivity analysis.

I. Ablation Studies

  • Joint HTRP and FAF achieve the best overall ablation results: 58.10% OA, 0.5391 Kappa, and 0.838 SSIM, with 783.8 ms inference time.The study reports this combination as balancing classification, structural preservation, and computational efficiency.Source: Table III · p. 11, §I · p. 11
  • HTRP alone raises OA from 55.39% to 57.08%, Kappa from 0.5097 to 0.5283, and SSIM from 0.805 to 0.833 versus baseline.The authors also report reduced inference time and note a slight SSIM disadvantage relative to FAF.Source: §I · p. 11
  • FAF alone yields 57.25% OA, 0.5306 Kappa, and 0.857 SSIM—the highest SSIM among the tested ablation configurations.Its inference time is 1149.4 ms, a slight increase relative to baseline.Source: §I · p. 11

J. Parameter Sensitivity Analysis

  • Increasing Top-K initially improves OA and Kappa but steadily increases inference time; K = 7 gives the highest OA and Kappa at reasonable inference time.Beyond K = 7, performance gains diminish relative to added computational cost.Source: Figure 7 · p. 12, 1) Top-K · p. 12
  • Replacing the Transformer decoder with a Linear decoder lowers OA from 58.10% to 57.41% and Kappa from 0.5391 to 0.5220.The authors attribute the Transformer decoder’s advantage to global modeling that refines contextual representations.Source: 2) Decoder Usage · p. 12, 1) Top-K · p. 12

2) Decoder Usage:

The paper compares BERT, GPT-2, and LLaMA3 on classification and inference efficiency, finding different trade-offs among the evaluated models. It concludes that RadioLLM combines HPTR and FAF for varied cognitive-radio tasks and conditions.

3) LLM Choice:

  • LLaMA3 has the highest OA and Kappa, GPT-2 the fastest inference, and BERT the lowest classification performance and longest inference time among the evaluated models.The comparison reports OA, Kappa, and inference time as its measures.Source: Table IV · p. 12, 3) LLM Choice · p. 12
  • The authors recommend selecting an LLM according to application constraints and priorities, balancing accuracy against computational efficiency.Source: 3) LLM Choice · p. 12

V. CONCLUSION

  • The conclusion describes RadioLLM’s HPTR as combining hardware and semantic software prompts, while FAF captures fine-grained high-frequency features for complex signal environments.The authors report experiments across tasks involving noisy conditions, imbalanced datasets, and intricate signal patterns.Source: §V · p. 12

Related papers

Loading 2501.17888v3…