Title: BERT4beam: Leveraging Large AI Models for Generalized Beamforming Optimization

Abstract:

Artificial intelligence (AI) is widely expected to serve as a critical driver for the next generation of sixth-generation (6G) wireless networks. Nevertheless, current studies involving large AI models in this domain predominantly concentrate on fine-tuning pre-trained large language models (LLMs) for particular applications. In contrast, this study explores a large-scale AI architecture specifically engineered for beamforming optimization, designed to adapt to and generalize across a wide array of tasks characterized by different system utilities and operational scales.

We introduce BERT4beam, a novel framework grounded in bidirectional encoder representations from transformers (BERT). This approach recasts beamforming optimization as a sequence learning problem at the token level. The methodology involves tokenizing channel state information, building the BERT structure, and implementing strategies for both task-specific pre-training and fine-tuning. Utilizing this foundation, we present two distinct BERT-based methodologies for beamforming optimization: one tailored for single-task scenarios and another for multi-task environments.

Both proposed solutions demonstrate robust generalization capabilities across varying user numbers. The single-task approach exhibits flexibility by adapting to changes in system utilities and antenna setups through the reconfiguration of the BERT model’s input and output modules. Meanwhile, the multi-task variant, referred to as UBERT, achieves direct generalization to diverse tasks by employing a more granular tokenization strategy. Comprehensive simulation results confirm that these two approaches deliver performance close to the theoretical optimum, surpassing existing AI models in various beamforming optimization contexts and highlighting their exceptional adaptability and generalization potential.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC