Introduction
Mobile Augmented Reality (MAR) applications demand high-precision pose estimation to synchronize virtual overlays with the physical environment. To mitigate device power consumption, these computationally intensive tasks are frequently offloaded to edge servers. However, supporting MAR in 6G networks faces two primary challenges: domain silos between network controllers and service providers, and the inherent volatility of user movement. Network controllers often lack visibility into application-specific rendering logic, while service providers are unaware of real-time network constraints. This paper proposes an LLM-powered "Digital Agent" to bridge this gap, enabling personalized, Quality of Experience (QoE)-aware resource management.
The Agent-Driven Service Provisioning Framework
The core of the proposed solution is a cross-layer architecture that facilitates communication between the network and service domains. By using a tool-calling framework, the agent enables personalized resource management that adapts to individual user movements while keeping sensitive application data private.
As shown in the architecture above, the framework consists of three primary components:
- Digital Agent (DA): Utilizing a Large Language Model (LLM) core, the DA serves as an intelligent intermediary. It processes natural language prompts from the network controller and translates them into structured commands for specialized tools.
- Service Function Toolkit (SFT) and MCP: The agent interacts with tools via the Model Context Protocol (MCP). A key tool is the Future QoE Prediction Tool, which integrates the service provider's pose prediction model with a Kalman Filter for adaptive bias correction, allowing for accurate real-time predictions despite user movement variability.
- Donor-Receiver Resource Algorithm: This algorithm optimizes bandwidth allocation by identifying "donor" users with surplus resources and reallocating them to "receiver" users experiencing declining QoE.
Experimental Methodology
The framework was evaluated using a real-world 6DoF pose dataset featuring head movement trajectories from 40 subjects viewing a volumetric video titled "Longdres." This dataset captures the non-stationary nature of human behavior in AR environments, providing a rigorous testbed for the agent's predictive capabilities.
Key Results and Discussion
The proposed agent-driven approach was benchmarked against a Base Transformer (BT) trained on general data and a Personalized Transformer (PT) trained on individual history.
The results indicate that the proposed method significantly outperforms the benchmarks in both Mean Squared Error (MSE) and User Category Accuracy. Unlike the Personalized Transformer, which often fails when encountering novel user behaviors, the agent-driven approach maintains stability by leveraging the underlying physics and rendering logic of the AR application.
Furthermore, the system demonstrated high resource efficiency. In a five-user simulation, the framework successfully redistributed bandwidth from stable users to those with dropping QoE, enhancing overall group immersion without increasing total spectrum usage.
Conclusion and Future Work
This research demonstrates that LLMs can serve as effective reasoning bridges in complex network environments, preserving privacy by sharing QoE insights rather than raw data. Future work will address the latency of the LLM reasoning loop to meet the millisecond requirements of 6G and explore scaling the architecture for high-density user environments. Moving toward "intent-driven" interaction will further enable autonomous negotiation between networks and service agents.
Summary updated