The paper Large Language Model Evaluation Via Multi AI Agents: Preliminary results introduces a novel way of assessing Large Language Models (LLMs). Using a multi-agent AI model, the researchers propose a comprehensive evaluation framework for various LLMs.

• The approach includes code retrieval from LLMs based on a common description by several AI agents.
• A verification agent evaluates the generated code against benchmarks like HumanEval.
• Preliminary results suggest GPT-3.5 Turbo leads in performance among the LLMs tested.

This research is central to understanding the effective deployment of LLMs in diverse environments. It provides a structural method to identify the strengths and weaknesses of different LLMs, thereby contributing to the responsible development and application of AI technologies. The potential for this kind of evaluation extends to addressing societal impacts and minimizing risks associated with LLM deployment.