LLM-Enabled Fusion for Distributed Emitter Perception

Abstract

Emitter spectrum cognition is a critical capability for electromagnetic environment awareness, where heterogeneous sensing observations must be integrated to support reliable emitter state estimation and signal behavior understanding. This paper investigates the potential of large language models (LLMs) for AI-based emitter spectrum cognition. Specifically, we explore whether pre-trained LLMs can leverage their inherent reasoning and contextual modeling capabilities to fuse heterogeneous emitter-related spectrum sensing information, including sequential state estimates, signal observations, and time-varying environmental cues. To bridge the gap between numerical sensing outputs and the semantic representation space of LLMs, we introduce a modality-alignment mechanism that transforms sequential emitter-spectrum state estimates into structured token embeddings compatible with the latent space of the LLM. Furthermore, a system-as-prompt (SaP) structure is designed to guide the LLM in understanding emitter spectrum dynamics, cognition objectives, and fusion requirements through taskspecific instructions. The proposed LLM-based emitter spectrum cognition framework enables robust information integration under time-varying electromagnetic conditions and demonstrates strong generalization capabilities across different sensing scenarios. Experimental results indicate that LLMs can serve as effective reasoning-enhanced fusion engines for emitter spectrum cognition in complex electromagnetic environments.