Core ML Runtime Flips Image Labels

This is the kind of paper that should make model teams uncomfortable in a useful way: benchmark wins in PyTorch do not guarantee the same behavior once the model hits an edge runtime. The nasty part is that the mismatch can show up without gradients, internals, or fancy attack tooling.

• –The risk looks infrastructure-level, not architecture-specific, because the drift comes from shared runtime behavior across common building blocks like Conv2D.
• –Core ML’s FP16 GPU path stands out as especially brittle in this write-up, which makes device/runtime selection part of the threat model.
• –Deployment validation should include cross-runtime regression tests, not just accuracy checks on the training framework.
• –The article’s 20% mismatch rate is preliminary, but it is high enough to justify treating runtime parity as a release gate.
• –For vision and edge-AI teams, this is a reminder that “same model” does not mean “same system.”