AI Can Now Read the Hidden History in Software Code

A new exploratory study tested whether large language models can recover architectural design decisions from source code commits—important because these decisions are rarely documented and stay hidden inside code changes. Using four LLMs with zero-shot and few-shot prompts on 30 developer-written decisions, the researchers found strong alignment scores overall, with few-shot prompting improving results. But the outputs were often too long, focused too much on implementation details, and missed the underlying rationale. The authors see this as a promising step toward architecture-aware LLM systems and automated architectural knowledge management—not a finished solution.