Integrating Pre-Trained Language Model with Physical Layer Communications

TL;DR

This work tackles on-device AI communication by integrating a pre-trained language model with a realistic PHY stack, enabling end-to-end transmission of semantic content over wireless channels. It introduces a VQ-VAE-based codebook to compress the AI-Src-Enc output and uses encoder-decoder transformers pre-initialized from BART to improve generalization, all trained with channel noise via $L_{CE}$ and $L_{VQVAE}$ losses. The framework is validated in a link-level 5G-NR–like environment with 3GPP CDL channels, Polar coding, QAM, OFDM, and MIMO, achieving up to 50% transmission-size reduction and EbN0 gains up to several dB while maintaining semantic fidelity, including out-of-distribution generalization. Limitations include the absence of MAC/HARQ/ARQ considerations; future work may extend to higher-layer protocols, refined noise-tuning, and improved PHY-function integration to further enhance reliability and efficiency.

Abstract

The burgeoning field of on-device AI communication, where devices exchange information directly through embedded foundation models, such as language models (LMs), requires robust, efficient, and generalizable communication frameworks. However, integrating these frameworks with existing wireless systems and effectively managing noise and bit errors pose significant challenges. In this work, we introduce a practical ondevice AI communication framework, integrated with physical layer (PHY) communication functions, demonstrated through its performance on a link-level simulator. Our framework incorporates end-to-end training with channel noise to enhance resilience, incorporates vector quantized variational autoencoders (VQ-VAE) for efficient and robust communication, and utilizes pre-trained encoder-decoder transformers for improved generalization capabilities. Simulations, across various communication scenarios, reveal that our framework achieves a 50% reduction in transmission size while demonstrating substantial generalization ability and noise robustness under standardized 3GPP channel models.

Figures (7)

• Figure 1: Overview of semantic communication systems and on-device AI/LM communication system.
• Figure 2: Framework overview of on-device AI/LM PHY communication systems integrated with pre-trained language model. This framework incorporates a link-level simulator to realistically emulate bit-level transmission within PHY communication systems. At the transmission (TX) end, the symbol stream $\vb* s$ undergoes AI-Src-Enc $S_{\mathrm{\mathbf{E}}}(\cdot)$, vector quantization $Q_{\mathrm{\mathbf{E}}}(\cdot; Z)$, and channel encoder $C_{\mathrm{\mathbf{E}}}(\cdot)$ to produce $\vb* x$. Conversely, at the receiver (RX) end, the received signal $y$ is channel-decoded, vector-dequantized, and semantically decoded to recover the symbol $\vb*{\hat{s}}$. We evaluate the system in three different criteria: lexical similarity, semantic similarity, and compression rate.
• Figure 3: Architecture of AI-Source-Encoder (AI-Src-Enc) and Quantizer
• Figure 4: Performance comparison with existing frameworks under different EbN0 [dB] (Channel = Rayleigh)
• Figure 5: Impact of noise-tuning (w/o Noise-Tuning vs. w/ Noise-Tuning). The CDL-A channel assesses the in-distribution performance, whereas the CDL-B, -C, and -D evaluate the out-of-distribution performance.