GRF-LLM: Environment-Aware Wireless Channel Modeling via LLM-Guided 3D Gaussians

Houding Zhang, Tianxiang Xu, Chang LIU, Xinyu Li

Keywords: wireless radiation field reconstruction, 3D Gaussian splatting, channel prediction

Abstract: We present GRF-LLM, a transformative framework for environment-aware wireless channel modeling that synergizes 3D Gaussian splatting with large language models (LLMs). Traditional wireless channel modeling approaches face critical limitations: probabilistic models lack spatial granularity for complex multi-antenna systems, while deterministic methods like ray tracing require precise environmental priors that are often unavailable or inaccurate in practice. Our proposed method addresses these challenges by introducing LLM-guided initialization and optimization of 3D Gaussian primitives that dynamically encode both signal propagation characteristics and material-dependent attenuation effects. The framework leverages LLMs' semantic understanding capabilities to analyze environmental scenes and predict material properties, enabling intelligent placement and parameterization of virtual transmitters represented as 3D Gaussians. These primitives are processed through a differentiable rendering pipeline that reconstructs wireless radiation fields with unprecedented efficiency and accuracy. The integration of LLM guidance enables automatic adaptation to diverse environmental conditions without requiring extensive manual parameter tuning or detailed material property databases. Through comprehensive experimental validation on real-world datasets, GRF-LLM demonstrates superior performance compared to existing methods, achieving a 3.82 dB improvement in downlink CSI prediction accuracy while maintaining real-time rendering capabilities. Our approach establishes a new paradigm for AI-enhanced propagation modeling, with significant implications for 6G network optimization and digital twin applications in wireless communications.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 10111