Ask Before You Diagnose: Safe-Psych, a Sequential Evaluation Benchmark for LLMs in Psychiatry

Short summary

Safe-Psych is a sequential benchmark evaluating how LLMs handle evolving diagnostic uncertainty in clinical psychiatry, using over 1,000 real-world psychiatric notes with psychiatrist-derived action labels (DIAGNOSE, CLARIFY, ABSTAIN). Findings reveal that even strong models exhibit over 60% under-abstention, frequently diagnose before sufficient evidence is available, and rarely seek clarification unless explicitly prompted. Safety-aware prompting shifts errors toward excessive abstention rather than improving calibration. The benchmark is released to support LLM safety research in healthcare.

• •Safe-Psych benchmarks LLMs on handling incomplete psychiatric diagnostic information across 1,000+ real clinical notes
• •Even strong models show >60% under-abstention and frequently diagnose prematurely without sufficient evidence
• •Safety-aware prompting reduces premature commitment but shifts errors toward excessive abstention rather than fixing calibration

Generated with AI, which can make mistakes.