Creative Writing by LLM Agents to Smoothly Reflect Human Intentions

Abstract

Recent studies on LLM (Large Language Model) agents focus on automating task resolution to minimize human intervention, thereby reducing effort. However, in tasks involving creative production activities, such as drafting service proposals, there is a challenge in sufficiently reflecting human intent. This study proposes a method for semi-automatically integrating concise human instructions into multi-agent task resolution processes. The method aims to enhance complex task design based on instruction prompts through reinforcement learning for automatic optimization, maximizing the reflection of human intent. By aligning outputs with human intent, the proposed approach fosters creativity in task resolution. This paper provides a detailed explanation of the proposed method and its evaluation framework.