New static analysis catches ML semantic bugs before training

A team of researchers led by Willem Meijer at Linköping University has introduced a novel static analysis technique that detects semantic faults in machine learning code during development, not after training. The approach, detailed in a paper accepted at ICSE-NIER 2026, tackles common issues like feeding unscaled data into scale-sensitive models or applying the wrong preprocessing pipeline. By leveraging combined data and control flow analysis alongside API contracts, the method can reason about data characteristics without executing the code—essentially catching bugs at write time instead of after hours of training.

The tool was validated on a sample of real-world ML notebooks, and the authors report it successfully identifies faults that require a data-aware perspective—faults that traditional static analyzers miss because they ignore data semantics. With ML models growing more complex and training costs skyrocketing, catching such errors early could save significant time, compute, and money. The paper highlights that this technique is especially valuable for production pipelines where data transformations are often opaque and error-prone.