EvoOR-Agent: Building an Adaptive and Interpretable Agent Optimization Framework Using Co-Evolution

Automation in operations research (OR) faces bottlenecks from manually designed workflows, and existing systems struggle to adapt to the diversity and complexity of OR problems. EvoOR-Agent represents agent architectures as AOE networks and applies co-evolutionary algorithms, outperforming zero-shot LLMs, fixed-pipeline agents, and existing evolutionary frameworks in heterogeneous OR benchmarks, achieving dual breakthroughs in performance and interpretability.

Operations research (OR) supports modern business operations, but converting OR problems into solvable models requires highly specialized knowledge. Existing OR automation systems rely on manually designed fixed reasoning-execution workflows, which have two major limitations: first, complex multi-stage coordination (needing to handle dependencies across problem interpretation, modeling, solver selection, etc.); second, inability to cope with the diversity of reasoning paths (multiple modeling and solving strategies exist for the same problem).

EvoOR-Agent treats agent architectures and reasoning trajectories as evolvable objects and adopts a co-evolution mechanism. The core innovation is the AOE network representation: nodes represent decision points/states, edges represent activities/reasoning steps, and paths represent complete reasoning trajectories, making the workflow structure explicitly visible. The framework maintains two evolutionary objects simultaneously: the architecture graph (defining workflow topology) and a population of reasoning individuals (specific path instances under architectural constraints), enabling co-optimization at both the architecture and instance levels.

EvoOR-Agent uses three evolutionary operators: 1. Graph-mediated path conditional recombination (structure-aware, ensuring recombination validity); 2. Multi-granularity semantic mutation (coarse/medium/fine-grained adjustments, balancing macro exploration and micro optimization); 3. Elitist population update (retaining optimal individuals, maintaining diversity). Additionally, the knowledge base auxiliary module injects prior knowledge (modeling/solving experience from similar cases) during initialization and guides mutation directions to accelerate convergence.

In heterogeneous OR benchmarks (linear programming, integer programming, combinatorial optimization, etc.), EvoOR-Agent performs excellently: outperforming zero-shot LLMs (improving success rate via structured reasoning), fixed-pipeline agents (adaptive architecture handles complex variants), and representative evolutionary frameworks (OR knowledge base enhances targeting). Case studies show that the explicit representation of AOE networks enhances interpretability, allowing tracing of the reasoning path of solutions and comparison of path merits.

Ablation experiments show: 1. Performance drops significantly when architecture evolution is disabled, indicating that workflow structure evolution is key to adapting to diverse OR problems; 2. When AOE graph structure is disabled, recombination validity decreases and convergence slows down—graph structure provides semantic guidance; 3. When knowledge base assistance is disabled, initial performance is poor and convergence is slow—the knowledge base provides prior guidance.

EvoOR-Agent points the way for OR automation: from manual design to automatic evolution, unification of performance and interpretability, and effective integration of domain knowledge. Limitations include computational overhead, problem scale constraints, and knowledge base dependency. Future research directions: introducing approximate evaluation to accelerate evolution, exploring distributed strategies, and expanding to fields like scientific discovery and engineering design.