Zero-shot performance analysis of large language models in sumrate maximization

Area of Science:

• Computer Science
• Artificial Intelligence
• Telecommunications Engineering

Background:

• Large language models (LLMs) are increasingly versatile, impacting diverse fields including natural language processing.
• In networking, LLMs offer potential for resource optimization and sharing, particularly for complex tasks like sumrate maximization.
• Traditional algorithms for sumrate maximization can be difficult to implement and understand.

Purpose of the Study:

• To evaluate the feasibility of using LLMs for sumrate maximization in networking without prior algorithmic knowledge.
• To assess the performance of LLMs in a zero-shot setting for network resource optimization.
• To compare LLM performance against state-of-the-art approaches for sumrate maximization.

Main Methods:

• A zero-shot analysis was conducted on LLMs to assess their suitability for sumrate maximization.
• Experiments involved varying the number of cellular users and device-to-device (D2D) pairs.
• The efficiency of LLMs was measured against established methods for sumrate maximization.

Main Results:

• The GPT model achieved a maximum average efficiency of approximately 58% for sumrate maximization.
• Performance varied based on the combinations of cellular users and D2D pairs tested.
• Some LLM variants demonstrated limitations with numerical and structural data in their current state.

Conclusions:

• LLMs, particularly GPT, show promise for simplifying sumrate maximization in networking.
• Further research and parameter fine-tuning are necessary for optimal LLM application to numerical network data.
• LLMs can potentially reduce the complexity associated with implementing advanced networking algorithms.