Cluster-Aware pipeline cuts auto software test generation from weeks to hours

A team of six researchers (Hazem Ayman, Menna Sedik, Kareem Mostafa, Mahmoud Soliman, Samer Saber, and Ibrahim Habib) published a new method for automated test specification generation in automotive software. Their paper, now on arXiv, targets the scalability bottleneck where manual test creation for Automotive SPICE SWE.6 compliance consumes weeks of engineering effort when projects reach thousands of requirements. Standard LLM approaches fail because they either process requirements individually (missing dependencies) or feed entire corpora (exceeding context windows), leading to incomplete coverage and redundant tests.

The proposed pipeline operates in three stages. First, it embeds requirements using sentence transformers, reduces dimensionality with UMAP, and clusters them via HDBSCAN with automatic cluster size selection optimized by a combined Silhouette and Calinski-Harabasz score. Second, a multi-level map-reduce summarization algorithm distills each cluster into concise descriptions while preserving quantitative thresholds and safety integrity levels. Third, it generates dual-level test specifications: individual requirement verification and cluster-level integration tests that verify cross-requirement feature behavior. A nearby-cluster context mechanism provides bounded cross-feature awareness during each LLM call, and Retrieval-Augmented Generation grounds all outputs in ISO 26262 and ASPICE standards. Evaluation on real automotive datasets shows improved integration test coverage and scaling to thousands of requirements.