The industrial adoption of executable visual workflows, prized for their reliability, is currently hobbled by manual engineering. Developers spend considerable time designing, prompting, and iterating on these complex systems, a process ripe for automation.
Bridging the Gap to Agentic Workflow Generation
To address this, researchers introduce the Chat2Workflow benchmark, a novel dataset comprising real-world business workflows designed for direct deployment on platforms like Dify and Coze. This benchmark serves as a critical tool to investigate the potential of large language models (LLMs) in automating the multi-round interaction required for workflow creation. The goal is to move beyond manual construction towards more autonomous systems.
The LLM Frontier: Intent vs. Execution
Experimental results reveal a persistent challenge: while state-of-the-art LLMs can grasp high-level user intent, they falter in generating workflows that are consistently correct, stable, and executable. This gap is particularly pronounced when dealing with complex or evolving requirements. The proposed agentic framework shows promise, achieving up to a 5.34% resolve rate gain in handling recurrent execution errors, yet a significant real-world gap remains, underscoring the need for continued advancement in industrial-grade automation.