Supervised models used for code analysis and optimization become brittle when the deployment environment drifts - a new CPU microarchitecture, runtime library, or workload pattern shifts the test distribution away from the training data. The Prom library restores robustness at deployment without retraining from scratch: it uses distribution-free statistical assessment (conformal prediction) to flag test samples likely to be mispredicted, then relabels a small fraction (up to 5%) of those flagged samples and folds them back in through incremental learning. Across 13 models and 5 code tasks it identifies about 96% of mispredictions and recovers near training-time accuracy.
Software-engineering / AI primary-research paper (Huanting Wang, Patrick Lenihan, Zheng Wang; University of Leeds; CGO 2025). Supervised ML for code analysis and optimization is brittle: minor change…