[3J5-GS-5] Agents:

[3J5-GS-5-04] A Planning Method for LLM Agents through Optimization of Top-Down Procedures

〇Natsuki Murakami¹, Tomoyuki Kagaya¹, Akira Kinose¹ (1. Panasonic Connect Co., Ltd)

Keywords:LLM Agent, Planning, Efficiency, Interpretability

Due to the remarkable achievements of Large Language Models (LLMs), research on LLM agents has become increasingly active. However, the inference methods of LLMs face several challenges, including a massive increase in the exploration space for complex tasks, inefficient planning, lack of consistency in reasoning, and insufficient explainability.
This study proposes an LLM agent that employs a planning framework inspired by cognitive psychology and logic to mimic human problem-solving abilities. Specifically, the agent organizes complex tasks into simpler sub-tasks by breaking them down to the most basic level.
The results of experiments conducted using Minecraft demonstrate that the proposed method facilitates the integration with low-level action generation. Furthermore, the improved explainability indicates the potential for the agent to collaborate effectively with humans.