Bridging 6G IoT and AI: LLM-Based Efficient Approach for Physical Layer's Optimization Tasks

TL;DR

This work tackles real-time, retraining-free optimization of physical-layer tasks in 6G IoT under energy-harvesting constraints by introducing the PE-RTFV framework, which uses two edge LLMs (an optimizer and an agent) guided by IoT feedback. The method relies on structured prompts and semantic guidance, enabling iterative refinement toward solutions that maximize mutual information $I(X;Y)$ while satisfying energy-harvesting requirements $P_H$. Experiments on a SWIPT-enabled constellation-design problem show that PE-RTFV can reach near-genetic-algorithm performance within a few iterations, even with minimal feedback such as 1–2 bits, and without explicit EH models. The results suggest that prompt-engineered, feedback-driven LLM control can provide scalable, lightweight optimization for complex physical-layer tasks in resource-constrained 6G IoT networks, offering practical impact for wireless-powered deployments.

Abstract

This paper investigates the role of large language models (LLMs) in sixth-generation (6G) Internet of Things (IoT) networks and proposes a prompt-engineering-based real-time feedback and verification (PE-RTFV) framework that perform physical-layer's optimization tasks through an iteratively process. By leveraging the naturally available closed-loop feedback inherent in wireless communication systems, PE-RTFV enables real-time physical-layer optimization without requiring model retraining. The proposed framework employs an optimization LLM (O-LLM) to generate task-specific structured prompts, which are provided to an agent LLM (A-LLM) to produce task-specific solutions. Utilizing real-time system feedback, the O-LLM iteratively refines the prompts to guide the A-LLM toward improved solutions in a gradient-descent-like optimization process. We test PE-RTFV approach on wireless-powered IoT testbed case study on user-goal-driven constellation design through semantically solving rate-energy (RE)-region optimization problem which demonstrates that PE-RTFV achieves near-genetic-algorithm performance within only a few iterations, validating its effectiveness for complex physical-layer optimization tasks in resource-constrained IoT networks.

Figures (6)

• Figure 1: System model.
• Figure 2: Overview of the PE-RTFV-based framework for physical-layer's optimization tasks, highlighting the roles of the structured prompt components (e.g., reward, feedback and semantic guidance) in generating the initial solution (points on the sampled objective-function curve) and adaptively selecting the optimization direction and learning rate in a gradient-descent-like iterative process.
• Figure 3: Constellations generated for the constellation design task using the PE-RTFV approach for users $U_1, U_2, \ldots, U_4$. Each user $U_i$ aims to maximize the achievable rate while ensuring $P_H \geq P_i^{\text{th}}$. The results are compared with constellations generated using the GA. The green $AQAM_{\text{GA}}$ constellations are obtained using the GA following mehmood2025asymmetric. The constellations $AQAM^{\text{FFB}}_{\text{LLM}}$ (red), $AQAM^{\text{2BFB}}_{\text{LLM}}$ (blue), and $AQAM^{\text{1BFB}}_{\text{LLM}}$ (pink) are generated using the full-feedback, 2-bit-feedback, and 1-bit-feedback PE-RTFV approaches, respectively. The PE-RTFV framework is executed for a maximum of 15 rounds, for each $P_i^{\text{th}}$ and each feedback type to generate these constellations.
• Figure 4: The performance comparison between constellations designed by the RT-FV PE approach and the baseline APSK and GA based approach. Simulation parameters are: M = 16, $\rho = 0.5$ for sub figure a) and b), and $\rho = [0,1]$ for fig c), N = $10^5$.
• Figure 5: The PE-RTFV based approach for device's goal specific constellation design.