Explicit-Constrained Single Agent for Enhanced Task-Solving in LLMs

Yi Zheng, Haibin Huang, Chongyang Ma, Luozhou Wang, Liuxiaoqiang, zhangyuan, Kanle Shi, Ying-Cong Chen, Pengfei Wan, Di ZHANG

Keywords: Agent, LLM

Abstract: In this study, we introduce the Explicitly Constrained Agent (EC-Agent), a novel approach designed to enhance the task-solving capabilities of Large Language Models (LLMs). Unlike existing multi-agent systems that depend on agents evaluating tasks from different perspectives, EC-Agent explicitly imposes task-oriented constraints for LLMs. Our observations are two-fold: first, assigning agents to sub-tasks with defined responsibilities implicitly sets constraints; second, these multi-agent systems often struggle with accurately assigning agents to sub-tasks, leading to overlapping duties and potential misguidance. In contrast, our single-agent system, driven by explicit methods and constraints, provides LLMs with detailed prompts, resulting in more precise responses. EC-Agent consists of two stages: a Reasoning Stage and a Summary Stage. 1) In the Reasoning Stage, three modules are proposed: Explicit Method, Explicit Constraint, and Execution. Specifically, LLMs utilize the Explicit Method and Constraint modules to analyze the task type and specific rules, generating multiple suitable methods and constraints. Subsequently, the Execution module combines these methods and constraints to produce and output possible solutions. 2) In the Summary Stage, LLMs evaluate the multiple reasoning processes and results from the previous step. They rectify any inconsistencies, summarize the information, and output the final result. Experimental results demonstrate that EC-Agent outperforms previous methods across a variety of tasks.

Primary Area: foundation or frontier models, including LLMs

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: 1894