LLM-Enhanced Pilot-Free OTFS Transceiver for High-Mobility Communications

Abstract

The advent of 6G wireless networks demands highly reliable communication in extreme mobility scenarios while maintaining stringent spectral efficiency. Conventional orthogonal time frequency space (OTFS) systems rely on pilot-assisted channel estimation, which incurs significant overhead and limits throughput. In this paper, we propose a novel pilot-free OTFS transceiver empowered by large language models (LLMs) to overcome these challenges. Our end-to-end (E2E) framework integrates artificial intelligence (AI)-driven constellation geometric shaping (GS) at the transmitter with an LLM-enhanced neural receiver (LLM-Rx) that performs implicit channel estimation and data detection directly in the delay-Doppler (DD) domain. A dual-branch receiver architecture combines convolutional feature extraction with a pretrained GPT-2 backbone, enabling context-aware joint estimation and detection. Extensive simulations demonstrate that the proposed design achieves 14.3% higher spectral efficiency and significantly lower bit error rate (BER) than pilot-assisted linear transceivers and other existing schemes.