.png){kind=icon}
Return
Cooperation makes better: Multi-agent interactive evolution framework in LLM4OR
.png){kind=icon}
PRE
AI
DOI：10.1016/j.eswa.2026.133754.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
This paper proposes a meta-level prompt optimization framework, Multi-Agent Interaction Prompt Optimization (MAIPO), to address two persistent limitations in Large Language Model for Operation Research (LLM4OR): limited generalization across heterogeneous tasks and insufficient integration of domain expertise. MAIPO comprises two core components. First, we abstract the multi-candidates iterative dialogue mechanism with domain expertise as a structural Partially Observable Markov Decision Process (POMDP) for reward-driven prompt evolution. Second, we integrate cooperative game theory into the multi-agent collaboration framework to quantify and allocate agent contributions. These two components jointly enhance the professionalism and effectiveness of prompt optimization, while preserving robustness in constrained computational settings. Extensive experiments across multiple benchmark datasets and representative baseline methods demonstrate that MAIPO consistently achieves substantial performance improvements peaking at 15.7%. Also, we illustrate that clearly defining the agent's role and task is a more cost-effective optimization strategy. Data and Code are available in https://github.com/ChengxiSHE/MAIPO.git.
Keywords:
Multi-agent framework
Agentic evolution
Markov decision process
Large language models for operation research
Journal
IF:
7.5
Papers:
3.0W
Citations:
10.2W
Organization
Cited Papers
No cited papers available