TCLNet: A Hybrid Transformer-CNN Framework Leveraging Language Models as Lossless Compressors for CSI Feedback

TL;DR

TCLNet tackles the high CSI feedback overhead in FDD massive MIMO by coupling a parallel hybrid Transformer–CNN lossy compressor with a LM–FM lossless compressor that adaptively routes symbols for entropy coding. The framework delivers superior rate–distortion–complexity trade-offs and demonstrates substantial reconstruction gains versus state-of-the-art baselines on real and simulated datasets, while also exploring zero-shot and prompt-engineered LLMs as universal lossless CSI compressors. Key findings include strong robustness to channel estimation errors, favorable RDC curves across quantization settings, and the potential of LLMs to approach entropy bounds in lossless coding. The work provides a practical, scalable approach for high-dimensional CSI feedback with configurable complexity, paving the way for deployment in large-scale MIMO systems and offering promising avenues for leveraging LLMs in physical-layer tasks.

Abstract

In frequency division duplexing (FDD) massive multiple-input multiple-output (MIMO) systems, downlink channel state information (CSI) plays a crucial role in achieving high spectrum and energy efficiency. However, the CSI feedback overhead becomes a major bottleneck as the number of antennas increases. Although existing deep learning-based CSI compression methods have shown great potential, they still face limitations in capturing both local and global features of CSI, thereby limiting achievable compression efficiency. To address these issues, we propose TCLNet, a unified CSI compression framework that integrates a hybrid Transformer-CNN architecture for lossy compression with a hybrid language model (LM) and factorized model (FM) design for lossless compression. The lossy module jointly exploits local features and global context, while the lossless module adaptively switches between context-aware coding and parallel coding to optimize the rate-distortion-complexity (RDC) trade-off. Extensive experiments on both real-world and simulated datasets demonstrate that the proposed TCLNet outperforms existing approaches in terms of reconstruction accuracy and transmission efficiency, achieving up to a 5 dB performance gain across diverse scenarios. Moreover, we show that large language models (LLMs) can be leveraged as zero-shot CSI lossless compressors via carefully designed prompts.

Figures (10)

• Figure 1: Illustration of CSI feedback in Massive MIMO systems. In FDD systems, the lack of reciprocity between uplink and downlink CSI makes direct estimation of downlink CSI challenging.
• Figure 2: Illustration of the proposed TCLNet, which consists of a lossy encoder based on a Transformer-CNN hybrid architecture and a lossless encoder based on LM and FM. The lossy encoder is used to compress the amount of CSI, while the lossless encoder is used for estimating the information entropy and encoding.
• Figure 3: Illustration of the proposed hybrid transformer-CNN-based lossy compressor, which is composed of both CNN and swin-transformer, focusing on extracting local and global features from the CSI, respectively.
• Figure 4: Illustration of the proposed LM- and FM-based coding scheme, where the LM encodes latent symbols with statistical dependencies, while the FM encodes symbols that are relatively independent.
• Figure 5: Illustration of the proposed LLM-based lossless compressor, where ASCII tokenization and in-context learning are employed along with an LLM and an arithmetic encoder.