Port-LLM: A Port Prediction Method for Fluid Antenna based on Large Language Models

Yali Zhang, Haifan Yin, Weidong Li, Emil Björnson, Mérouane DebbahIEEE Trans. Communications 2025

In the world of wireless communications, movement is often a "curse." When a smartphone moves quickly—say, in a car or a high-speed train—the wireless channel changes so rapidly that the base station struggles to keep up, leading to dropped packets and sluggish speeds. Traditionally, we’ve treated antennas as fixed points. But what if the antenna could "flow" to a better position in real-time?

A new paper titled "Port-LLM: A Port Prediction Method for Fluid Antenna based on Large Language Models" proposes a radical solution: using the sequence-prediction power of Large Language Models (LLMs) to predict where a mobile antenna should move to keep your connection rock-solid.

TL;DR

Researchers have developed Port-LLM, the first framework to use pre-trained Large Language Models (like GPT-2) to predict the optimal position—or "port"—for a Fluid Antenna on a moving device. By leveraging LLMs' superior sequence modeling, the system achieves over 99% prediction accuracy and millisecond-level latency, effectively turning a chaotic, time-varying channel into a stable one.

What Problem Does This Paper Address?

The authors target two main hurdles in modern wireless tech:

  1. The Space Constraint: On small devices (UE), we can’t fit many traditional fixed antennas. This limits "spatial degrees of freedom"—our ability to exploit different signal paths.
  2. The Curse of Mobility: When a device moves, the Channel State Information (CSI) becomes obsolete almost instantly.

Fluid Antennas (FA) are a novel technology where an antenna can switch between many tiny "ports" within a small area. If we can predict which port will have the best signal a few milliseconds into the future, we can move the antenna there and maintain a "stationary" channel experience even while traveling at 150 km/h. However, predicting this is incredibly difficult because the relationship between a port's location and its signal quality is highly non-linear.

Methods

The researchers introduced two models: Port-LLM and Prompt-Port-LLM. Both follow a two-step logic: first, predict the future "channel table" (the signal quality at every possible port), and then select the port that most closely matches a stable reference signal.

Key Innovations:

  • LLM Backbone: Instead of building a model from scratch, they used a pre-trained GPT-2 architecture. They argue that the "knowledge" LLMs gain from processing language sequences translates well to processing wireless signal sequences.
  • LoRA Fine-Tuning: To make the model efficient, they used Low-Rank Adaptation (LoRA). This allows them to "nudge" the massive LLM toward wireless tasks by only updating a tiny fraction of its parameters, keeping the rest frozen.
  • Prompt-Port-LLM: This version uses "Dynamic Prompts"—textual descriptions of the task and real-time statistics of the data—to guide the LLM, much like how you might give a specific persona to ChatGPT to get better results.
  • Dimensionality Reduction: Since channel data is complex, they designed a specialized Encoder Block (using CNNs and downsampling) to compress the wireless data into a format the LLM can digest.

Datasets

The models were evaluated using a simulated environment based on:

  • 3GPP CDL-D Channel Model: A standard for simulating realistic 5G/6G environments.
  • High-Frequency/High-Speed: Operating at 39 GHz with UE speeds ranging from 90 km/h to 150 km/h.
  • Scale: A dataset of 54,300 samples collected from 10 different UEs with varying orientations.
  • Configuration: A 2D fluid antenna moving area of $10\lambda \times 20\lambda$ with 5,000 possible ports ($100 \times 50$).

Key Insights and Ideas

  • LLMs vs. Traditional NNs: Standard recurrent networks (RNNs, LSTMs) often suffer from "gradient explosion" when dealing with complex wireless sequences. The Transformer-based LLM architecture proved much more stable and robust.
  • The "Tolerance Range": The authors found that the model doesn't need to be 100% perfect. Because ports are so dense, as long as the model picks a port within a certain "tolerance range" of the optimal one, the communication performance remains excellent.
  • Zero-Shot Generalization: Remarkably, a model trained on a simple Single-Input Single-Output (SISO) setup worked effectively in Multiple-Input Single-Output (MISO) scenarios without any retraining.

Key Results

  • High Accuracy: Port-LLM achieved 98.18% accuracy, while Prompt-Port-LLM reached 99.45%.
  • Superior Stability: The Normalized Mean Square Error (NMSE) of the predicted ports converged to below -30 dB, significantly outperforming traditional methods like Vec Prony and MPMP.
  • Millisecond Inference: Port-LLM takes only 4.72 ms to make a prediction, making it fast enough for real-time hardware implementation.
  • Spectral Efficiency: In multi-user scenarios, the LLM-based port selection provided much higher data throughput (Spectral Efficiency) compared to existing non-AI benchmarks.

Limitations and Future Work

  • Computational Overhead: While inference is fast, LLMs are still "heavier" than simple linear filters. Implementing this on low-power mobile chips remains a challenge.
  • Simulation vs. Reality: The study relies on 3GPP simulations. Real-world hardware for fluid antennas (like liquid metal or pixel-based antennas) may introduce mechanical switching delays not fully accounted for here.
  • Multi-User Complexity: Future work could explore how port prediction changes when multiple users are competing for the same spatial resources, potentially leading to interference that the current model hasn't fully mapped.

Paper Citation: Y. Zhang et al., "Port-LLM: A Port Prediction Method for Fluid Antenna based on Large Language Models," arXiv:2502.09857v2, 2025.

Summary updated

Abstract

The objective of this study is to address the mobility challenges faced by user equipment (UE) through the implementation of fluid antenna (FA) on the UE side. This approach aims to maintain the time-varying channel in a relatively stable state by strategically relocating the FA to an appropriate port. To the best of our knowledge, this paper introduces, for the first time, the application of large language models (LLMs) in the prediction of FA ports, presenting a novel model termed Port-LLM. Our proposed method for predicting the moving port of the FA is a two-step prediction method. To enhance the learning efficacy of our proposed Port-LLM model, we integrate low-rank adaptation (LoRA) fine-tuning technology. Additionally, to further exploit the natural language processing capabilities of pre-trained LLMs, we propose a framework named Prompt-Port-LLM, which is constructed upon the Port-LLM architecture and incorporates prompt fine-tuning techniques along with a specialized prompt encoder module. The simulation results show that our proposed models all exhibit strong generalization ability and robustness under different numbers of base station antennas and medium-to-high mobility speeds of UE. In comparison to existing methods, the performance of the port predicted by our models demonstrates superior efficacy. Moreover, both of our proposed models achieve millimeter-level inference speed.
BibTeX
@article{zhang2025port,
  author = {Yali Zhang and Haifan Yin and Weidong Li and Emil Björnson and Mérouane Debbah},
  title = {{Port-LLM: A Port Prediction Method for Fluid Antenna based on Large Language Models}},
  journal = {IEEE Trans. Communications},
  year = {2025},
}
Figure