arXiv · 2611.18574
Predictive Test Optimization Without Historical Failure Data
Abstract
As software systems grow in size and complexity, their test suites become increasingly expensive to execute, delaying feedback and wasting resources. State-of-the-art predictive test optimization methods use machine learning to select tests relevant to a change, typically trained on historical failures. This depends on failure histories and is susceptible to label noise from flaky tests. We introduce the novel method of co-evolution labeling for predictive test optimization, deriving test relevance from tests and code changing together in the version history. These data are widely available and less noisy. We train and evaluate models with failure and co-evolution labels on three large projects---from industry and open source; each with thousands of tests and more than 1M LOC. Co-evolution labeling nearly matches the failure detection capabilities of failure-based models, with an overall Average Percentage of Faults Detected of 0.88, while being more resistant to label noise and requiring no test result history. Co-evolution labeling offers a practical, immediately deployable path to faster pipelines and reliable predictive test optimization.
Explore related subjects
Keep this discovery
Explore connections, maps & timelines
Maximilian Jungwirth, Raphael Nömmer, Andreas Stahlbauer, Sven Apel, Gordon Fraser. 2026-07-27. Predictive Test Optimization Without Historical Failure Data. https://searcharxiv.com/abs/2611.18574
Citations retain the original authors and DOI and include this site’s record URL. If this discovery service contributed to your research, you can also cite Searcharxiv by Indiciae: BibTeX · RIS.