Divide-and-Conquer Is What LLM-Based Multi-Agent System Need

ACL ARR 2025 July Submission432 Authors

Abstract:

Large language model (LLM) based multi-agent systems offer promising capabilities in social simulation and complex task solving, yet face key challenges in system design, generalizability, and scalability.
We introduce \method, a novel framework featuring:
(1) a fully parallel divide-and-conquer architecture for efficient task decomposition and distributed processing;
(2) an adaptive collaboration engine that dynamically selects heterogeneous LLMs and interaction strategies;
(3) agent organization optimization for effective problem breakdown.
Experiments show that \method achieves state-of-the-art results across several benchmarks, with substantial improvements on tasks such as GSM8K, AIME, and HumanEval, especially as task complexity increases.
Our results demonstrate that \method enables the construction of robust and general-purpose LLM multi-agent systems, excelling in complex reasoning scenarios.

Paper Type: Long

Research Area: NLP Applications

Research Area Keywords: Language Modeling

Contribution Types: NLP engineering experiment

Languages Studied: english

Submission Number: 432