EvoGraph: Hybrid Directed Graph Evolution Toward Software 3.0
7 hours ago
- #software-evolution
- #automated-code-modernization
- #graph-based-framework
- Introduces EvoGraph, a framework for software systems to evolve source code, build pipelines, documentation, and tickets autonomously.
- Represents artifacts in a typed directed graph, uses learned mutation operators from specialized small language models (SLMs), and selects survivors via multi-objective fitness.
- Achieves 83% security vulnerability fixes, 93% functional equivalence in COBOL-to-Java translation, and maintains documentation freshness within two minutes.
- Shows 40% latency reduction and sevenfold drop in feature lead time compared to baselines; extends to modernization of .NET, Lisp, CGI, ColdFusion, Python, and C with 82-96% semantic equivalence.
- Reduces computational costs by 90% versus large language models and addresses legacy modernization issues like implicit contracts and performance preservation.