On Using Large Language Models to Generate Plans
Abstract: Automated planning is concerned with developing efficient algorithms to generate plans or sequences of actions to achieve a specific goal in a given environment. Emerging Large Language Models (LLMs) can answer questions, write high-quality programming code, and predict protein folding, showcasing their versatility in solving various tasks beyond language-based problems. This paper explores if and how LLMs can also be used for automated planning given the diverse ways LLMs are modeled and trained. To do so, we seek to answer four key questions. Firstly, we want to understand the effectiveness of different LLM architectures for plan generation. Secondly, we aim to identify which pre-training data (general purpose vs code specific) effectively facilitates plan generation. Thirdly, we investigate whether fine-tuning or prompting is a more effective approach for plan generation. Finally, we explore whether LLMs are capable of plan generalization. By answering these questions, the study seeks to shed light on the capabilities of LLMs in solving complex planning problems and provide insights into the most effective approaches for using LLMs in this context.
Submission Length: Regular submission (no more than 12 pages of main content)
Changes Since Last Submission: 1. In the Introduction, paragraph 4 has been removed in the latest draft to avoid redundancy.
2. The Background and Related Work section has been modified as per the feedback.
3. In Appendix A.1 Frequently Asked Questions, the major contributions are removed.
4. Table 4 has been shortened to fit in one page.
5. The figure on token lengths for PDDL and compact representation has been moved to Appendix A.6.
6. Appendix A.7 has been added with the NL prompt examples for zero-shot, one-shot, and CoT prompting techniques.
7. All additions made in the revised paper are highlighted in $\textcolor{blue}{blue}$.
Assigned Action Editor: ~Hector_Palacios1
Submission Number: 2401
Loading