On October 24, 2026, a technical report titled “Protegrity-PII Technical Report” was published on the arXiv platform (ID 2610.30518v1), officially introducing Protegrity-PII—a novel decision model specifically designed to identify personally identifiable information (PII) in both structured and unstructured text.
The model achieved a typed recall of 0.915 and a precision of 0.907 on the researchers’ newly constructed benchmark, PII-SCOPE. Its inference efficiency is 0.49 seconds per million characters, with a unit processing cost of $0.0005 per million characters. The testing environment consisted of a single NVIDIA RTX 6000 GPU, and batch inference was employed.
The report notes that Protegrity-PII demonstrates superior generalization performance in out-of-distribution settings compared to several current mainstream models and datasets, partly attributable to the fine-grained, broad‑coverage entity type classification scheme employed during training. The PII‑SCOPE benchmark itself was also introduced for the first time in this work and has not yet appeared in other public literature.