arXiv:2611.21622v1 Announce Type: new Abstract: Background: Crisis helplines are essential to mental health systems but are often understaffed and underfunded. LLMs could support counselors, but existing benchmarks do not capture the escalation, risk assessment, and de-escalation central to crisis conversations, while privacy restrictions limit use of real transcripts. Methods: We present CrisisBench, a privacy-preserving suite for single- and multi-turn crisis counseling evaluation. It includes 1,085 single-turn tasks from 200 de-identified conversations and 1,185 patient LLMs derived from hotline metadata. Patient models and a 23-item LLM judge were calibrated against expert counselors. Results: We evaluated five GPT models plus draft-then-edit and retrieval-augmented generation. All systems underperformed human counselors on excessive empathy and questioning in single-turn evaluation. The rubric achieved human-human Krippendorff's $\alpha$ = 0.51 and LLM-human $\alpha$ = 0.54 (Pearson r = 0.80). In multi-turn evaluation, retrieval-augmented generation scored highest overall (0.961), with all systems within 0.058. Microskills and style/tone were near ceiling, while productivity and safety/risk accounted for most between-system variation. Counselors rated patient LLMs realistic (86%) and useful (92%). Conclusion: CrisisBench enables scalable, privacy-preserving evaluation of crisis counseling. Current systems perform strongly on core structural competencies but still diverge from experts on important response-level failure modes. We release the rubrics, judge prompts, and evaluation tasks.
CrisisBench: an evaluation suite for crisis counseling AI
Researchers released CrisisBench, a privacy-preserving evaluation suite for crisis counseling AI built from 1,085 single-turn tasks drawn from 200 de-identified conversations and 1,185 patient LLMs derived from hotline metadata, with a 23-item LLM judge calibrated against expert counselors. In testing, five GPT models plus draft-then-edit and retrieval-augmented generation all underperformed human counselors on excessive empathy and questioning in single-turn evaluation, while in multi-turn evaluation retrieval-augmented generation scored highest overall (0.961) with all systems within 0.058. The rubric reached human-human Krippendorff's alpha of 0.51 and LLM-human alpha of 0.54 (Pearson r = 0.80), and counselors rated the patient LLMs realistic (86%) and useful (92%); the authors released the rubrics, judge prompts, and evaluation tasks.
Calibrate an LLM judge against expert human raters and report agreement (CrisisBench hit alpha 0.54, r = 0.80) before trusting it to score crisis or safety-sensitive responses.
more in playbook: Evals and testing for LLM appsRun your AI side-project on zahid.host
EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.