arXiv cs.LG
· Papers
Decodable but Not Detectable: A Leakage Fingerprint for Near-OOD Benchmarks
arXiv:2607.19393v1 Announce Type: new Abstract: While auditing a perturbation-based OOD detector on a document benchmark, we recorded an AUROC of 0.326 -- well below the 0.5 chance level. The cause is a benchmark leak: the designated "OOD" class is one the model was trained on, so its examples sit inside the in-distrib