New AI System Automates Discovery of Hidden Dependencies Between Software Requirements

Tracking how different software requirements relate to and depend on one another remains a persistent bottleneck in software engineering. As systems grow more complex, the sheer volume of interconnected requirements makes manual dependency mapping increasingly impractical—yet this groundwork shapes everything from design decisions to testing strategy.

A new research paper (arXiv:2602.22456) proposes an automated solution: LEREDD, an approach built on large language models that can systematically uncover requirement dependencies from natural language descriptions.

The Challenge#

Requirement dependency detection is traditionally done by hand or ignored entirely. Modern software systems amplify the difficulty: requirements are numerous, tightly coupled, and often written in vague or evolving language. This ambiguity makes it hard for conventional rule-based tools to parse relationships reliably, which is why the task has largely remained manual.

How LEREDD Works#

LEREDD combines two LLM techniques: Retrieval-Augmented Generation (RAG) and In-Context Learning (ICL). Rather than relying on generic zero-shot prompting, the approach contextualizes decisions by retrieving similar requirement pairs and learning from examples during inference. This helps the model distinguish between truly dependent requirements and those that merely happen to be related in topic.

Performance and Scale#

Testing against two established baselines, LEREDD achieved:

• 93% accuracy in classifying dependent vs. non-dependent requirements
• F1 score of 0.84 overall, with 0.96 for non-dependent cases
• 94.87%–105.41% relative gains in F1 scores when detecting specific dependency types (e.g., "Requires" relationships)

Researchers also released an annotated dataset of 813 requirement pairs across three systems to enable reproducible research and benchmarking in the field.