EvoGraphCoder: An Evolutionary Graph-Reasoning Framework for Self-Adaptive Software Engineering
Karthik Ramamurthy, Rahul Kumar Konduru, Nurmyrat Amanmadov · IEEE Access · 2026
Modern software evolves rapidly, accumulates architectural debt, and develops cross-module dependencies that complicate reliable maintenance. We presentEvoGraphCoder, an evolutionary graph-reasoning framework for self-adaptive software engineering. EvoGraphCoder represents source code, tests, commit history, dependencies, performance signals, and review feedback as a relational software graph. It combinesAdaptive Evolutionary Code Reasoning(AECR) with a multi-agent design consisting of theInnovator, Critic, andHistorianto generate, evaluate, and refine candidate repairs over multiple cycles rather than emitting a single patch. The framework further introducesEvoGraph Memoryfor persistent cross-release learning andSelf-Reflexive Validationfor explainable pre-merge verification. Experiments on repository-level repair benchmarks show that EvoGraphCoder improves patch quality and robustness over strong baselines, while maintaining positive improvement across repeated repair cycles. These results suggest that graph-driven evolutionary reasoning with persistent memory offers a practical path toward reliable and explainable AI-assisted software maintenance.