← All benchmarks

PSEBench

Health & Medicine · 2026-06-03

PSEBench, from Emory University, Scale AI, Mayo Clinic and Vanderbilt University Medical Center, tests patient safety event triage under Minnesota's 29 Reportable Adverse Health Events. Given an incident narrative with the statute and its guidance, the model may put factual questions to a GPT-5.2 information provider that holds the case's facts, over at most 10 turns, and then answers Reportable, Non-Reportable or Uncertain with the targeted clause, the legal basis and a rationale. Its 5,074 cases are generated from 146 clause cards with verified by-construction labels: 3,455 complete reports, 1,362 that withhold decision-critical facts the model must ask for, and 257 regulatory gray-zone cases that should be escalated as Uncertain. The score is the paper's M1, three-way verdict accuracy over all 5,074 cases from one run per model; the clause, evidence-citation, boundary-condition, information-seeking and uncertainty metrics (M2 to M8) are not imported.

Top models (higher is better)