Prompt-Enabled Large AI Models for CSI Feedback

TL;DR

This paper investigates the interpretability and generalization of AI-based CSI feedback and proposes a prompt-enabled large AI model (LAM) that leverages environmental knowledge as prompts to improve CSI reconstruction. It demonstrates that AI-based CSI feedback benefits from both strong fitting capabilities and utilization of environment-specific information, and that prompts fed to the decoder can further boost reconstruction quality without online training for new scenarios. The approach uses a transformer-based autoencoder with six-block encoders/decoders, processes CSI in the angular-delay domain, and incorporates mean CSI magnitude as a prompt to embed environmental knowledge. Across extensive QuaDRiGa- and MATLAB-generated datasets, the prompt-enabled LAM outperforms small-scale models, generalizes well to unseen scenarios, and reduces data collection overhead, indicating practical potential for scalable, adaptive CSI feedback in diverse environments.

Abstract

Artificial intelligence (AI) has emerged as a promising tool for channel state information (CSI) feedback. While recent research primarily focuses on improving feedback accuracy on a specific dataset through novel architectures, the underlying mechanism of AI-based CSI feedback remains unclear. This study explores the mechanism through analyzing performance across diverse datasets, with findings suggesting that superior feedback performance stems from AI models' strong fitting capabilities and their ability to leverage environmental knowledge. Building on these findings, we propose a prompt enabled large AI model (LAM) for CSI feedback. The LAM employs powerful transformer blocks and is trained on extensive datasets from various scenarios. Meanwhile, the channel distribution (environmental knowledge) -- represented as the mean of channel magnitude in the angular-delay domain -- is incorporated as a prompt within the decoder to further enhance reconstruction quality. Simulation results confirm that the proposed prompt-enabled LAM significantly improves feedback accuracy and generalization performance while reducing data collection requirements in new scenarios.

Figures (11)

• Figure 1: Detailed architecture of the FC-based CSI feedback NN, where the encoder at the user and the decoder at the BS compress and reconstruct channel vector ($\bf h$) using FC layers, respectively.
• Figure 2: NMSE performance versus the number of Res-FC blocks ($N$) for varying cluster numbers ($L$) and codeword lengths ($N_{\rm c}$). Regardless of the $N_{\rm c}$ and $L$, increasing the number of Res-FC blocks leads to improved feedback accuracy.
• Figure 3: NMSE performance of AI-based CSI feedback versus the user range radius ($R$). The solid line represents the NMSE performance in the specific scenario generated in Section \ref{['observationSimulation']}, while the dashed line represents the NMSE performance in the general scenario, where channels are randomly generated using MATLAB based on spatial channel model. The cluster number ($L$), Res-FC block number ($N$), and the codeword length ($N_{\rm c}$) are set to 4, 5, and 6, respectively.
• Figure 4: Illustration of LAMs without and with prompts. In this illustration, the application of prompts, specifically in specifying question types, directs LAMs toward the intended outcomes.
• Figure 5: Illustration of the general framework of prompt-enabled LAMs for CSI feedback, including offline training and online inference.