daic

This paper audits benchmark evaluation in clinical-interview depression detection through four complementary probes across five datasets, finding that standard evaluation protocols may overestimate model performance and that leaderboard rankings lack stability.