DistressBench: Evaluating Multi-Dimensional Crisis Support in Large Language Models

Key Findings

• — The best model (Meta Muse Spark 1.2) satisfied 88.4% of weighted criteria and the median model 71.4%, against 98.7% for clinician reference responses scored on the same rubrics; Llama 4 Maverick scored lowest at 41.2%.
• — Pooled across all 25 models, in 35.3% of conversations where a model met every criterion for recognising the crisis it then failed to route the user toward human help; the leading model missed the handoff in 15.9% of the conversations it recognised.
• — Median dimension scores were 94.5% for avoiding harmful content and 98.2% for avoiding moralisation, against 46.0% for de-escalation (best model 74.6%), 70.3% for actionable resources and 8.3% for safety disclaimers (best model 84.6%).
• — Multi-turn tasks scored below single-turn tasks (68.5% against 74.2%).
• — Claude Fable 5 returned an empty final turn on 162 of 2,154 task-attempts (7.5%); blanks were most frequent in the escalation-pattern (34.2%), trauma-triggered distress (20.7%) and uncontrolled-anger (15.3%) subcategories and rare for suicidal ideation and suicide planning (0.8% each). The authors attribute the blanks to the served API endpoint and score them as zero.
• — When shown a failed criterion, models usually recognised the compliant reply but did not treat it as the better message (85.0% of forfeited weight), which the authors read as a behavioural choice rather than a capability limit.

Methodology Notes

718 tasks (299 single-turn, 419 multi-turn of 3 to 5 user turns) across 23 suicide and self-harm subcategories, labelled by age group (under or over 18), harm intent, progression stage, turn structure and signal directness. Rubrics (6,901 criteria, 6 to 20 per task, weights 1 to 20) were model-drafted from clinician notes and then rewritten by the authoring clinician (88.5% of drafted criteria reworded, 1,731 added). Replies were scored by a frozen three-judge LLM panel calibrated on 75 stratified tasks against three clinician raters per criterion; three attempts per model; 95% bootstrap confidence intervals. English only; the corpus is simulated, not real user conversations. The public half (359 tasks) was posted to Hugging Face on 2026-10-12 under CC BY 4.0 (click-through gated) with the 75-task calibration set, per-judge labels for the 25 evaluated models, the grader and long-horizon user trajectories; the private half is held out and scored on request. The live leaderboard ranks more models than the paper (33 on 2026-10-13); figures in this record are from the paper. Includes a 20-transcript long-horizon red-team annex reported separately. Authors are Scale AI staff; the paper is a company preprint, not peer-reviewed. Paper page dated 2026-10-09 (PDF created 2026-10-06); verified from the Scale Labs paper page, blog post, leaderboard page and the 39-page PDF.

Sources