LLM-powered system cuts cost of analyzing requirement changes by 95%

Software requirements evolve constantly as projects progress, but pinpointing which other requirements are affected by each change remains a labor-intensive and error-prone task for engineering teams. A new study from researchers at arXiv introduces ProReFiCIA, an automated system leveraging large language models to tackle this problem efficiently.

The challenge of change impact analysis#

When a single requirement shifts in a software development project, the ripple effects can be significant—related requirements may become incompatible, redundant, or incomplete. Manually tracing these dependencies across large requirement documents is both time-consuming and prone to oversight. Engineers working within tight budgets often cannot afford to perform exhaustive impact analysis, which can lead to overlooked dependencies and downstream problems if changes are not properly tracked.

How ProReFiCIA works#

The approach uses LLMs to automatically detect which requirements are affected whenever specifications change. The researchers tested multiple LLMs and prompt strategies customized for this specific task. In their evaluation on an unseen industrial dataset, the best-performing combination achieved 85.7% recall—meaning it correctly identified roughly 9 out of 10 impacted requirements.

Beyond accuracy, the system proved cost-effective: reviewers only needed to examine the AI-flagged requirements, which typically represented just 3.0% of the entire requirement set. This dramatic reduction in review scope makes the approach practical for real-world deployment.

Boosting accuracy with domain knowledge#

The researchers further improved performance by incorporating domain-specific context through retrieval-augmented generation (RAG). This enhancement pushed recall up to 95.8% while keeping the review burden modest at only 3.4% of total requirements. By grounding the LLM's analysis in organization-specific knowledge, the system becomes more attuned to subtle dependencies and business logic unique to each project.

Why it matters#

This work addresses a genuine bottleneck in software engineering: the gap between the increasing complexity of requirements and the finite resources available to manage them. Automating change impact analysis while preserving human oversight could help teams catch problems earlier, reduce manual effort, and ultimately improve software quality.