AutoManual: Generating Instruction Manuals by LLM Agents via Interactive Environmental Learning

基于大型语言模型（LLM）的代理已经在各个领域展示了自主完成任务的潜力，例如机器人、游戏和网络导航。然而，这些代理通常需要精心设计和专家提示才能解决特定领域的任务，这限制了它们的适应性。我们介绍了AutoManual，这是一个框架，使LLM代理能够通过交互自主建立对环境的理解并适应新环境。AutoManual将环境知识分成不同的规则，并通过两个代理以在线方式对其进行优化：1）规划者根据当前规则编写可操作的计划以与环境交互。2）构建者通过一个结构良好的规则系统更新规则，以方便在线规则管理和必要的细节保留。为了减少规则管理中的幻觉，我们引入了“案例条件提示”策略，用于构建者。最后，公式化代理将这些规则编译成全面的手册。自动生成的手册不仅可以提高适应性，还可以在可读性方面指导更小的LLM的规划。在仅有一个简单演示的情况下，AutoManual显著提高了任务成功率，在ALFWorld基准任务中，使用GPT-4-turbo和GPT-3.5-turbo分别达到了97.4％和86.2％。源代码将很快提供。

AutoManual: Autonomously Building an Interactive Agent's Manual from Environmental Knowledge

The AutoManual framework enables LLM agents to autonomously build their understanding through interaction and adapt to new environments by categorizing environmental knowledge into diverse rules and optimizing them in an online fashion by two agents: the Planner and the Builder. The self-generated manual can improve adaptability and guide the planning of smaller LLMs while being human-readable.

AutoManual significantly improves task success rates, achieving 97.4% with GPT-4-turbo and 86.2% with GPT-3.5-turbo on ALFWorld benchmark tasks. The framework introduces a case-conditioned prompting strategy for the Builder to mitigate hallucinations in managing rules. The source code will be available soon.

Related work includes recent research in the use of LLM agents for task completion across various domains, as well as work on rule-based systems for knowledge representation and management.