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

TL;DR

RadioLLM introduces a universal cognitive radio framework by embedding radio signal processing inside a pre-trained large language model. It combines Hybrid Prompt and Token Reprogramming (HPTR) to fuse expert knowledge with radio embeddings, and a Frequency Attuned Fusion (FAF) module to capture high-frequency signal details, supported by a lightweight decoder and LoRA-based fine-tuning. The approach excels across modulation classification and denoising tasks on diverse datasets, achieving state-of-the-art accuracy and robustness, especially at higher SNRs and in few-shot regimes. The results demonstrate strong cross-domain generalization, improved denoising fidelity (SSIM), and competitive inference efficiency, highlighting RadioLLM as a scalable, multi-task CRT solution with practical implications for spectrum management and wireless system optimization.

Abstract

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.

Figures (7)

• Figure 1: Comparison of existing CRT frameworks with our proposed approach.
• Figure 2: The model framework of RadioLLM. The input radio signal is preprocessed to generate signal embeddings $X_{s}$ (A). In the HPTR stage (B), $X_{s}$ is reprogrammed with semantic anchors $E'$, and top-K semantic anchors are selected as prefix prompt $P'_t$. The FAF stage (C) injects high-frequency features to address the transformer's low-pass filtering tendency. Finally, the enhanced embeddings and prefix prompts are fed into the LLM (A&D), which outputs denoised signals $O_s$ or classification results depending on the task.
• Figure 3: An illustrative example of the prompt template used in this study. The template incorporates domain-specific descriptions, statistical attributes, and task-specific instructions, which are color-coded for clarity. This unified prompt design enables the model to process heterogeneous signal information with rich contextual and structural cues.
• Figure 4: Performance evaluation of RadioLLM across multiple datasets and SNR levels. Each column corresponds to one dataset: RML2016a, RML16B, RML16C, RML22, and RML18A. (a)–(e) show the OA of RadioLLM compared to other models under varying few-shot sample settings. (f)–(j) depict the OA under different SNR conditions with the few-shot number fixed at 100. (k)–(o) present the confusion matrices of RadioLLM on the corresponding datasets, illustrating detailed classification results across modulation classes.
• Figure 5: Comparison of OA for different methods on RML16A, RML16B, and RML16C. “All” denotes variants pre-trained jointly on all three datasets. The results illustrate the effect of multi-domain pre-training on model adaptability and generalization.