ArchPipeline: A Two-phase Transfer-learning Framework for Software Architecture Comprehension and Evolution
Year: 2026
Type: article
Abstract: Software architecture evolution remains largely manual and error-prone.This paper presents ArchPipeline, a two-phase transfer-learning framework automating the comprehension and evolution of software architectures in ACME and AADL .The framework first fine-tunes a model to generate natural language descriptions from ADL (Architecture Description Language) specifications, then transfers the resulting weights to initialize an evolution model that produces structurally valid evolved ADL from natural language requests.Two domain-specific datasets are publicly available on Zenodo.The comprehension model achieves a composite score of 0.6152, while the evolution model reaches a composite score of 0.9431 on the full 120-instance test set.In a separate, sample-matched comparison on a 20-instance subset, it outperforms six one-shot LLM (Large Language Model) baselines (8B-120B active parameters) on all five evaluation metrics (EM, BLEU, ROUGE-L, METEOR, BERT-F1).An ablation study shows a modest but statistically significant benefit of two-phase initialization over direct fine-tuning on sequence-level metrics, demonstrating that domain-specific sequential fine-tuning on curated ADL data consistently surpasses one-shot prompting of general-purpose large language models.
Language: English
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7204649604