Evaluating Large Language Models in Crisis Detection: A Real-World Benchmark from Psychological Support Hotlines (2506.01329v1)
Abstract: Psychological support hotlines are critical for crisis intervention but face significant challenges due to rising demand. LLMs could support crisis assessments, yet their capabilities in emotionally sensitive contexts remain unclear. We introduce PsyCrisisBench, a benchmark of 540 annotated transcripts from the Hangzhou Psychological Assistance Hotline, assessing four tasks: mood status recognition, suicidal ideation detection, suicide plan identification, and risk assessment. We evaluated 64 LLMs across 15 families (e.g., GPT, Claude, Gemini, Llama, Qwen, DeepSeek) using zero-shot, few-shot, and fine-tuning paradigms. Performance was measured by F1-score, with statistical comparisons via Welch's t-tests. LLMs performed strongly on suicidal ideation detection (F1=0.880), suicide plan identification (F1=0.779), and risk assessment (F1=0.907), improved with few-shot and fine-tuning. Mood status recognition was more challenging (max F1=0.709), likely due to lost vocal cues and ambiguity. A fine-tuned 1.5B-parameter model (Qwen2.5-1.5B) surpassed larger models on mood and suicidal ideation. Open-source models like QwQ-32B performed comparably to closed-source on most tasks (p>0.3), though closed models retained an edge in mood detection (p=0.007). Performance scaled with size up to a point; quantization (AWQ) reduced GPU memory by 70% with minimal F1 degradation. LLMs show substantial promise in structured psychological crisis assessments, especially with fine-tuning. Mood recognition remains limited due to contextual complexity. The narrowing gap between open- and closed-source models, combined with efficient quantization, suggests feasible integration. PsyCrisisBench offers a robust evaluation framework to guide model development and ethical deployment in mental health.