RadioLLM-V2: Prototype-Augmented Large Language Models for Few-Shot Cognitive Radio Sensing

Abstract

Deep learning-based cognitive radio systems achieve strong results on modulation classification and SNR estimation, yet most models remain task-specific and depend on substantial labeled data. Large language models (LLMs) offer a more general interface, but applying them to radio signals requires bridging the gap between continuous I/Q sequences and discrete tokens. We propose RadioLLM-V2, an LLM-based framework for joint modulation classification and SNR estimation in few-shot settings. Distinct from traditional numerical position vectors, the proposed framework introduces a Semantic Positional Prompting (SPP) mechanism. SPP converts patch indices and sample-level time windows into language-driven positional cues, making patch-level temporal information more explicit in the LLM-compatible representation. The framework encodes I/Q samples with a dual-domain tokenizer and augments tokens with these semantic cues. Global signal statistics are converted into a compact physical prompt for prototype retrieval, and a prototype-guided Q-Former compresses the signal–prototype context into query tokens for a parameter-efficiently adapted LLM backbone. Experiments on five public benchmarks under the 100-shot protocol show that the proposed method achieves the best overall performance among the compared methods. On RML2016.10a, it improves classification accuracy from 58.33% to 59.29% and reduces SNR estimation RMSE from 3.12 to 2.74 compared with the original RadioLLM. It also attains 57.17% accuracy on the challenging RML2018 dataset, outperforming representative CNN/Transformer baselines.