Powering Cognitive Radio with AI

Abstract

Large Language Models (LLMs), particularly
transformer-based models, are increasingly used as
high-level agents in software systems. This paper
presents a proof of concept for applying LLMs to
GNU Radio by fine-tuning an open-source model
(Qwen Code 1.5B Instruct and others) on flowgraph
construction tasks. We introduce a modular training
pipeline built on Hugging Face libraries, along with
custom GNU Radio Companion (GRC) tracing tools that
capture flowgraph construction and runtime behavior.
This enables domain-specific dataset generation and
prototyping of LLM-driven signal processing agents.
While effective for orchestration, LLMs are poorly
suited to latency-sensitive and resource-constrained
environments. To address this, we propose Hebbian
Cellular Automata (HCA), a novel model that does
not rely on backpropagation. Inspired by local learning
rules and modulated feedback, HCA is fully recurrent,
synchronous, and well-suited for parallel, low-latency
systems. We show preliminary results on simple
pattern recognition tasks, positioning HCA as a
promising lightweight model. This work traces a
research trajectory from LLM-based orchestration
toward biologically inspired learning models better
aligned with real-time signal processing.