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

Shuai Chen, Yong Zu, Zhixi Feng, Shuyuan Yang, Mengchang Li, Yue Ma, Jun Liu, Qiukai Pan, Xinlei Zhang, Changjun Sun

Abstract

The increasing scarcity of spectrum resources and the rapid growth of wireless devices have made efficient management of radio networks a critical challenge. Cognitive Radio Technology (CRT), when integrated with deep learning (DL), offers promising solutions for tasks such as radio signal classification (RSC), signal denoising, and spectrum allocation. However, existing DL-based CRT frameworks are often task-specific and lack scalability to diverse real-world scenarios. Meanwhile, Large Language Models (LLMs) have demonstrated exceptional generalization capabilities across multiple domains, making them a potential candidate for advancing CRT technologies. In this paper, we introduce RadioLLM, a novel framework that incorporates Hybrid Prompt and Token Reprogramming (HPTR) and a Frequency Attuned Fusion (FAF) module to enhance LLMs for CRT tasks. HPTR enables the integration of radio signal features with expert knowledge, while FAF improves the modeling of high-frequency features critical for precise signal processing. These innovations allow RadioLLM to handle diverse CRT tasks, bridging the gap between LLMs and traditional signal processing methods. Extensive empirical studies on multiple benchmark datasets demonstrate that the proposed RadioLLM achieves superior performance over current baselines.

Introduction

Cognitive Radio Technology (CRT) is essential for managing our increasingly crowded wireless spectrum. While deep learning has improved tasks like Radio Signal Classification (RSC) and denoising, existing models are usually "one-trick ponies"—they are trained for a specific task or signal type and struggle to generalize to new environments.

Large Language Models (LLMs) offer incredible cross-domain generalization, but they aren't natively built for radio. Adapting them requires overcoming three main hurdles: the semantic gap between text and raw I/Q (In-phase/Quadrature) sequences, inefficient prompting that slows down inference, and frequency blindness, where the transformer backbone acts as a low-pass filter, missing the high-frequency phase shifts critical to radio signals.

The RadioLLM Framework

To address these challenges, the researchers introduced RadioLLM, which repurposes a pre-trained LLM (specifically GPT-2) to handle complex radio signal processing tasks. The framework utilizes three key innovations to bridge the gap between language and signals.

The RadioLLM architecture, featuring Hybrid Prompt and Token Reprogramming (HPTR) and Frequency Attuned Fusion (FAF).

Hybrid Prompt and Token Reprogramming (HPTR)

Instead of converting signals into wordy text descriptions, RadioLLM uses a multi-head cross-attention layer to map raw signal patches directly into the LLM’s embedding space. This "token reprogramming" allows the model to process the signal as if it were a sequence of words. To inject expert knowledge (e.g., "this is a radar signal"), the authors replace long, fixed text templates with "semantic anchors"—software-based embeddings retrieved from a joint semantic space. This approach improved inference speed by 31.85% while increasing accuracy, as LLMs process dense semantic embeddings more efficiently than natural language when dealing with non-textual data.

Frequency-Attuned Fusion (FAF)

Since Transformers often miss high-frequency details, the authors added a CNN-based FAF module. CNNs are naturally adept at detecting local variations. The FAF module extracts these high-frequency features and fuses them with the LLM’s global contextual understanding, ensuring that sharp signal details are preserved during processing.

Experimental Results

The model was evaluated on a comprehensive suite of seven public datasets, including the RadioML series (RML16A, B, C, and RML18A), RML22 for zero-shot testing, and real-world data from ADS-B (aircraft signals) and Wi-Fi.

Performance evaluation of RadioLLM across multiple datasets, highlighting few-shot accuracy and SNR robustness.

Key Performance Insights

  • Few-Shot Dominance: RadioLLM’s most significant advantage lies in low-data regimes. In 1-shot and 5-shot scenarios, it maintains a substantial accuracy lead over traditional models that require thousands of samples to converge.
  • Universal Representation: The model outperformed baselines on the RML22 dataset despite never seeing it during pre-training. This suggests that RadioLLM has learned a "universal" way to represent radio signals.
  • Noise Resilience: As shown in the SNR curves, the model maintains competitive performance in high-noise environments where other models' accuracy typically plummets.
  • Denoising Quality: In signal recovery tasks, it achieved a Structural Similarity Index (SSIM) of 0.838 to 0.893, significantly outperforming specialized denoising models like DNCNet.

Conclusion and Future Work

RadioLLM represents a significant step toward general-purpose artificial intelligence for wireless communications. By combining the reasoning power of LLMs with specialized signal-processing modules, the framework achieves state-of-the-art performance across diverse tasks. Future work will focus on refining the model's ability to distinguish between highly similar modulation schemes (like 16-QAM vs. 64-QAM) and scaling the architecture to larger backbones like LLaMA-3 to further enhance radio cognition.

Summary updated

Abstract

The increasing scarcity of spectrum resources and the rapid growth of wireless device have made efficient management of radio networks a critical challenge. Cognitive Radio Technology (CRT), when integrated with deep learning (DL), offers promising solutions for tasks such as radio signal classification (RSC), signal denoising, and spectrum allocation. However, existing DL-based CRT frameworks are often task-specific and lack scalability to diverse real-world scenarios. Meanwhile, Large Language Models (LLMs) have demonstrated exceptional generalization capabilities across multiple domains, making them a potential candidate for advancing CRT technologies. In this paper, we introduce RadioLLM, a novel framework that incorporates Hybrid Prompt and Token Reprogramming (HPTR) and a Frequency Attuned Fusion (FAF) module to enhance LLMs for CRT tasks. HPTR enables the integration of radio signal features with expert knowledge, while FAF improves the modeling of high-frequency features critical for precise signal processing. These innovations allow RadioLLM to handle diverse CRT tasks, bridging the gap between LLMs and traditional signal processing methods. Extensive empirical studies on multiple benchmark datasets demonstrate that the proposed RadioLLM achieves superior performance over current baselines.
BibTeX
@misc{chen2025radiollm,
  author = {Shuai Chen and Yong Zu and Zhixi Feng and Shuyuan Yang and Mengchang Li and Yue Ma and Jun Liu and Qiukai Pan and Xinlei Zhang and Changjun Sun},
  title = {{RadioLLM: Introducing Large Language Model into Cognitive Radio via Hybrid Prompt and Token Reprogrammings}},
  year = {2025},
  eprint = {2501.17888},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2501.17888},
}
Figure