Automatically Detecting Numerical Instability in Machine Learning Applications via Soft Assertions
This paper presents Soft Assertions, a technique for detecting and triggering numerical instability bugs in machine learning applications. The approach learns runtime guards around numerically sensitive operations and uses those guards to guide fuzzing toward inputs that expose instability.
The work targets silent ML failures where numerical issues corrupt predictions without producing obvious crashes, NaNs, or infinities.
BibTeX
@article{sharmin2025softassertions,
title = {Automatically Detecting Numerical Instability in Machine Learning Applications via Soft Assertions},
author = {Sharmin, Shaila and Zahid, Anwar Hossain and Bhattacharjee, Subhankar and Igwilo, Chiamaka and Kim, Miryung and Le, Wei},
journal = {Proceedings of the ACM on Software Engineering (PACMSE), issue FSE 2025},
year = {2025},
doi = {10.1145/3729394},
url = {https://arxiv.org/pdf/2504.15507}
}