AI Model Fixes Broken Software Requirements Links in Automotive Systems

Maintaining accurate connections between stakeholder needs and system requirements is a persistent challenge in automotive software engineering, with broken or missing links often leading to costly inconsistencies and compliance headaches. A new research paper from arXiv proposes TVR, a technique that harnesses large language models combined with retrieval-augmented generation (RAG) to both identify flawed traceability relationships and recover ones that have been lost.

The Problem#

In automotive development, engineers must manually map stakeholder requirements—what customers or regulators need—to technical system requirements that guide implementation. This manual process is error-prone: requirement changes sometimes fail to propagate correctly, and human mistakes during mapping create gaps. Most existing research has sidestepped this particular problem, focusing instead on other requirement relationships, and has rarely tested solutions against real industrial data where these links exist in practice. Additionally, automotive specifications frequently vary in how they're written, a inconsistency that trips up many machine-learning-based approaches.

How TVR Works#

Rather than relying on training data alone, TVR leverages the generalization power of large language models enhanced with retrieval-augmented generation—a technique that retrieves relevant context before generating responses. This design lets the system handle the natural variation in how automotive requirements are expressed without needing large volumes of labeled training examples. The approach both validates whether existing requirement links are correct and fills in missing ones.

Practical Impact#

Experimental results suggest TVR performs well in real-world automotive settings, offering a practical way to improve traceability in the kinds of complex systems where manual mapping becomes prohibitively difficult. The work points to a promising path for reducing the inconsistencies and maintenance burden that plague current automotive software development.