Title: Evolving Agent Structures and Transparent Reasoning for Automated Optimization

Abstract:

The integration of Large Language Models (LLMs) into operations research (OR) is currently hindered by its reliance on manually designed reasoning-execution workflows. Addressing intricate OR challenges demands dynamic coordination across several stages, including problem interpretation, mathematical modeling, solver selection, code generation, and iterative debugging. To overcome these constraints, we introduce EvoOR-Agent, a co-evolutionary framework designed for automated optimization. This system models agent workflows as Activity-on-Edge (AOE) networks, thereby clarifying the topology, execution dependencies, and alternative reasoning paths. Leveraging this representation, the framework sustains an architecture graph and advances a population of reasoning entities via graph-mediated, path-conditioned recombination, multi-granularity semantic mutation, and elitist population updates. Additionally, a knowledge-base-assisted experience-acquisition module integrates reusable OR practices into both initialization and semantic variation processes. Evaluations across diverse OR benchmarks demonstrate that our framework consistently outperforms zero-shot LLMs, static-pipeline OR agents, and prominent evolutionary agent frameworks. Further case studies and ablation analyses reveal that explicit architectural evolution and graph-based reasoning-trajectory search enhance both performance and structural interpretability. These findings indicate that viewing agent architectures and reasoning trajectories as evolvable entities offers a viable pathway to adaptive and interpretable automated optimization.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC