FlowAgent: a New Paradigm for Workflow Agent
Keywords: workflow, LLM-based agent, task-oriented dialog
Abstract: Combining workflows with large language models (LLMs) allows LLMs to follow specific procedures, thereby extending their application to more real-world scenarios. However, incorporating workflows often compromises the flexibility of LLMs. For example in the case of Task-Oriented Dialogue (TOD), workflow atomize the function of LLM while programmatically imposing restrictions on execution path making the dialogue obstructed and less flexible when facing out-of-workflow (OOW) queries. Prompt-based methods offer soft control but sometimes fail to ensure procedure compliance. This paper introduces a new agent paradigm to address this challenge. Specifically, we first propose a novel Procedure Description Language (PDL) that integrates the flexibility of natural language and the precision of code for workflow expression. Additionally, we present a comprehensive framework that enables LLM to handle OOW queries while keeping execution safe with a series of controllers for behavioral regulation. This includes pre-decision and post-decision methods, where the dependency relationships between workflow nodes are modeled as a Directed Acyclic Graph (DAG) to validate node transitions. Beyond the primary objective of compliance considered in previous work, we introduce a new approach to evaluate the agent's flexibility in OOW situations. Experiments on three datasets demonstrate that FlowAgent not only adheres well to workflows but also responds better to OOW queries, showcasing its flexibility. Furthermore, exploration on WikiHow data confirms that the PDL effectively represents broader formats of workflow, inspiring further research on workflow-based QA tasks.
Primary Area: applications to computer vision, audio, language, and other modalities
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 13875
Loading