Abstract
The rapid development of Large Language Models (LLMs) presents significant opportunities for 6G communications, particularly in network optimization and management through natural language interfaces. However, native LLMs face challenges such as a lack of private domain knowledge and limited reasoning capabilities. This paper proposes a multi-agent system (MAS) framework integrating retrieval, planning, memory, and evaluation to enhance LLMs for 6G. The system comprises three components: Multi-agent Data Retrieval (MDR), Multi-agent Collaborative Planning (MCP), and Multi-agent Evaluation and Reflexion (MER). A case study on semantic communication system design validates the effectiveness of this approach.
Introduction
The transition to 6G requires managing hyper-dynamic environments and diverse application needs like VR and autonomous driving. Traditional AI models are often too specialized and require extensive retraining. While LLMs offer universal reasoning, they struggle in 6G due to data privacy concerns, a lack of deep technical protocol knowledge, and a tendency to "hallucinate" during complex logical optimization tasks.
Proposed Framework: MDR, MCP, and MER
To bridge the gap between general-purpose LLMs and specialized 6G requirements, the authors introduce a multi-agent architecture. This framework moves beyond a single-agent interaction to a collaborative loop of specialized agents.
The framework consists of three core modules:
- Multi-agent Data Retrieval (MDR): Uses a Secure Agent to validate requests and Condensate/Inference Agents to extract technical knowledge from private vector databases, ensuring the system stays updated with the latest 6G standards without compromising privacy.
- Multi-agent Collaborative Planning (MCP): Employs Planning Agents to decompose high-level tasks into sub-task chains using techniques like Chain-of-Thought (CoT), ensuring logical consistency in complex designs.
- Multi-agent Evaluation and Reflexion (MER): Creates a self-correction loop where an Evaluation Agent scores outputs, and Reflexion/Refinement Agents provide feedback to iteratively improve solutions.
Case Study: Semantic Communication System Design
The system was tested by tasking it to design a Semantic Communication (SC) system for text transmission. Using a knowledge base of research papers and the Cornell Movie-Dialogs Corpus, the MAS autonomously generated Python code for the communication modules.
A key insight was the system's ability to evolve its own architecture. Initially proposing a Multi-Layer Perceptron (MLP), the MER loop identified that a Long Short-Term Memory (LSTM) network would better handle temporal dependencies in text, leading to a superior design without human intervention.
As shown in the results, the evaluative score increased significantly over four iterations. The final LSTM-based model achieved a BLEU score of 0.68 at 10 dB SNR, surpassing the user requirement of 0.6 while maintaining a strict parameter count under 2 million for edge compatibility.
Conclusion and Future Outlook
This research demonstrates a paradigm shift from AI as a simple classifier to AI as an autonomous system designer. By leveraging cognitive synergy among multiple agents, the framework mitigates LLM hallucinations and solves domain-specific 6G tasks. Future work will focus on reducing the computational cost and latency of these iterative loops to meet the millisecond-level response requirements of real-time 6G networks.
Summary updated