A new tool uses large language models to flag which software requirements secretly depend on each other - a job most teams still do by hand, if they do it at all.
Researchers built LEREDD, a system that pairs retrieval-augmented generation with in-context learning to scan natural language requirements and classify how they relate. Tested against two existing state-of-the-art baselines, LEREDD hit 93% accuracy distinguishing dependent from non-dependent requirement pairs, with an F1 score of 0.84 (0.96 on the easier non-dependent cases). It pulled further ahead on the harder task of naming the specific dependency type - for the "Requires" category alone, it beat the two baselines by 94.87% and 105.41% in F1 score. The team also released an annotated dataset of 813 requirement pairs from three different systems so the results can be checked.
Requirement dependencies determine what breaks when you change one spec, and a missed one is how integration bugs and scope creep sneak into large projects. Most teams skip this analysis by hand because it doesn't scale past a few hundred requirements, so an automated pass this accurate is a genuinely useful, if unglamorous, addition to the requirements engineering toolbox.
It won't replace a requirements engineer's judgment, but it might finally replace the spreadsheet.