Datasets used to train mental-health AI often carry bare labels: this text signals depression, this one doesn't, with no record of why. That makes the label hard to trust and makes it hard to build a model that can explain itself. The authors want annotations that show their work, tied explicitly to the clinical criteria in the DSM-5-TR.

Their framework pairs a language model with a human expert in a loop, moving through three steps: pulling candidate evidence from the text, checking it against each diagnostic criterion, then synthesizing a label with a severity rating. A two-part memory stores worked examples and reflections on expert corrections, so the system improves without retraining, and it exports the evidence, reasoning, and edit history so a person can audit it. They frame this as a tool for building better datasets, not for diagnosing anyone.

One honest caveat: the abstract reports only a pilot and leaves fuller evaluation to future work, so read the paper before drawing conclusions.