EvoOR-Agent Framework Co-evolves LLM Architectures for Automated Optimization Research

A new framework called EvoOR-Agent has been developed to improve the automation of operations research using large language models. It addresses the limitations of current LLMs that struggle with complex OR tasks, which usually rely on manually designed workflows. EvoOR-Agent represents workflows as activity-on-edge networks, making it easier to understand the structure, dependencies, and different reasoning paths. It includes a graph architecture and evolves reasoning agents through techniques like path-conditioned recombination, semantic mutations, and elite updates. There's also a knowledge-based module that integrates reusable OR strategies during both setup and variation. This research, identified as arXiv 2604.17708v1, focuses on automating tasks like problem interpretation and debugging.