LLMs Extract Architectural Design Decisions From Source Code Commits

An exploratory study shows Gemini 3 Pro, DeepSeek R1, Kimi K2, and Qwen3 can extract implicit architectural decisions from code commits. All tested models achieved a BERT-F1 score above 0.81, with few-shot prompting further improving alignment with human-written documentation.