Researchers posted DistressBench, a dataset of 718 English crisis conversations, on arXiv on 2026-10-12. Each conversation includes clinician‑written user turns and a rubric that rates seven dimensions of support. The benchmark scores models by the fraction of required care they deliver, not just whether they refuse to answer. Among 25 evaluated models, the top performer satisfied 88.4% of weighted criteria, while clinicians scored 98.7% on the same rubric.
Why it matters
The benchmark lets developers see where models fall short in real crisis help, especially in routing users to human assistance.
