📄 中文摘要
大型语言模型在复杂规划任务中面临挑战，尤其是在需要多步骤推理和状态跟踪时。本文提出了一种名为“知识模型提示”（KMP）的新方法，旨在通过将规划任务分解为可管理的子问题来增强LLM的规划能力。KMP的核心思想是利用LLM生成一个“知识模型”，该模型包含任务相关的状态表示、动作定义和转换函数。随后，LLM利用这个知识模型进行规划，通过迭代地应用动作并更新状态来探索解决方案路径。实验结果表明，KMP显著提高了LLM在多个规划基准测试中的性能，包括块世界、旅行商问题和物流规划。与传统的零样本或少样本提示方法相比，KMP能够更有效地处理复杂约束和长序列决策。此外，研究还发现，通过显式地构建和利用知识模型，LLM不仅能生成更准确的规划，还能提供更具解释性的推理过程。KMP的成功表明，将结构化知识引入LLM的提示机制是提升其在
📄 English Summary
Knowledge Model Prompting Increases LLM Performance on Planning Tasks
Large Language Models (LLMs) often struggle with complex planning tasks, particularly those requiring multi-step reasoning and state tracking. This paper introduces Knowledge Model Prompting (KMP), a novel approach designed to enhance LLM planning capabilities by decomposing planning problems into manageable sub-problems. The core idea of KMP is to leverage the LLM itself to generate a “knowledge model” comprising task-relevant state representations, action definitions, and transition functions. Subsequently, the LLM utilizes this generated knowledge model for planning, iteratively applying actions and updating states to explore solution paths. Experimental results demonstrate that KMP significantly improves LLM performance across various planning benchmarks, including Blocksworld, Traveling Salesperson Problem, and logistics planning. Compared to traditional zero-shot or few-shot prompting methods, KMP more effectively handles complex constraints and long-sequence decision-making. Furthermore, the study reveals that by explicitly constructing and utilizing a knowledge model, LLMs not only generate more accurate plans but also provide more interpretable reasoning processes. The success of KMP suggests that incorporating structured knowledge into LLM prompting mechanisms is a crucial avenue for boosting their performance in complex cognitive tasks.