Exécution de notre outil de synthèse sur un nouvel article

Résumé : Agentic Workflows are an important application form of Large Language Models (LLMs), but automatically generating executable Agentic Workflows from natural language descriptions is a significant challenge. Directly prompting LLMs to generate complex, logically sound, and executable workflows is often ineffective due to the complexities of task decomposition, node selection, parameter configuration, and inter-node dependency management. To address these challenges, we propose AutoFlowGen, a multi-agent framework that decomposes the workflow generation task into distinct stages: requirements analysis, structural planning, sketch drafting, code generation, and reflective optimization. Each stage is managed by specialized agents, which collaborate to enhance the quality and reliability of the generated workflows. Furthermore, to rigorously evaluate the performance of automated workflow generation systems, we introduce FinQA-Workflow, a novel benchmark dataset. This dataset follows an end-to-end evaluation philosophy and is specifically designed to test the quality of generated workflows in solving complex tasks, comprising 2000 question-answer pairs from the financial domain. Extensive experiments on FinQA-Workflow demonstrate that AutoFlowGen significantly improves the accuracy of workflows generated for solving complex financial problems, achieving an average accuracy improvement of 3.46% and up to 6.13% for specific models compared to direct prompting methods with various LLMs.