LLM4FB: A One-Sided CSI Feedback and Prediction Framework for Lightweight UEs via Large Language Models

Area of Science:

• Wireless communication systems
• Signal processing
• Artificial intelligence

Background:

• Massive MIMO enhances spectral efficiency but requires accurate channel state information (CSI).
• High antenna counts increase feedback overhead, and current deep learning methods burden user equipment (UE) computationally.
• Existing CSI feedback methods struggle with efficiency and UE resource constraints.

Purpose of the Study:

• To propose LLM4FB, a novel one-sided CSI feedback framework utilizing a pre-trained large language model (LLM).
• To reduce feedback overhead and computational complexity for UEs in Massive MIMO systems.
• To enhance CSI reconstruction accuracy and spectral efficiency with minimal UE resources.

Main Methods:

• The UE employs low-complexity linear projections for CSI compression.
• A pre-trained LLM at the base station (BS) reconstructs and predicts CSI.
• A multi-objective loss function optimizes Normalized Mean Square Error (NMSE) and Spectral Efficiency (SE).

Main Results:

• LLM4FB achieves high-precision CSI feedback across various compression ratios and mobility levels.
• The framework outperforms existing CSI feedback methods.
• Minimal fine-tuning of LLM parameters is required, reducing training costs.

Conclusions:

• LLM4FB offers a promising solution for CSI feedback in resource-constrained environments.
• The framework is suitable for next-generation wireless sensor networks and industrial IoT.
• LLM4FB effectively addresses the challenges of feedback overhead and UE computational burden.