Plan, Generate and Match: Scientific Workflow Recommendation with Large Language Models

Yang Gu, Jian Cao, Yuan Guo, Shiyou Qian, Wei Guan

Abstract: The recommendation of scientific workflows from public repositories that meet users’ natural language requirements is becoming increasingly essential in the scientific community. Nevertheless, existing methods that rely on direct text matching encounter difficulties when it comes to handling complex queries, which ultimately results in poor performance. Large language models (LLMs) have recently exhibited exceptional ability in planning and reasoning. We propose “Plan, Generate and Match” (PGM), a scientific workflow recommendation method leveraging LLMs. PGM consists of three stages: utilizing LLMs to conduct planning upon receiving a user query, generating a structured workflow specification guided by the solution steps, and using these plans and specifications to match with candidate workflows. By incorporating the planning mechanism, PGM leverages few-shot prompting to automatically generate well-considered steps for instructing the recommendation of reliable workflows. This method represents the first exploration of incorporating LLMs into the scientific workflow domain. Experimental results on real-world benchmarks demonstrate that PGM outperforms state-of-the-art methods with statistical significance, highlighting its immense potential in addressing complex requirements.