Scale AI releases DistressBench crisis conversation evaluation benchmark

1. 23:00
   Scale AI Research / SEALSelected

   Scale AI 发布 DistressBench，评测大模型在自杀与自伤危机对话中的支持能力

   Scale AI 的研究团队发布 DistressBench，用执业临床医生撰写的评分标准衡量大模型在自杀与自伤对话中是否真正提供了所需支持，而不是只看模型是否拒答。该基准包含 718 段由临床医生撰写的对话，覆盖 23 个自杀与自伤子类别，每段对话配有加权评分标准，共 6,901 条，涵盖危机支持的七个维度。在 25 个前沿模型上，表现最好的模型满足 88.4% 的加权标准，中位数模型为 71.4%，临床医生参考回答为 98.7%。最大差距出现在识别危机与采取行动之间：在模型完全识别出危机的对话中，仍有 35.3% 没有把用户引向人类帮助。模型在避免有害内容上中位数得分达 94.5%，但在缓和情绪上中位数仅 46.0%，最强模型为 74.6%，作者认为这一差距主要来自产品选择，因为至少已有一个部署中的模型表现明显更好。

2. Time not given
   Scale AI Research BlogSelected

   Scale AI releases DistressBench: a clinician-written benchmark for crisis help scenarios

   Scale AI releases DistressBench, which uses 718 English psychological crisis conversations written by practicing clinicians or crisis counselors to evaluate how well general-purpose chatbots provide help in suicide and self-harm scenarios, rather than merely whether they refuse. The benchmark covers 23 suicide and self-harm subcategories, with 299 single-turn and 419 multi-turn conversations. Each conversation is paired with 6 to 20 binary criteria written by clinicians and weighted by importance (1 to 20), totaling 6,901 criteria across seven dimensions: identification, empathetic engagement, de-escalation, actionable resources, disclaimers, moral judgment, and non-harmful responses. The score is the proportion of weighted criteria met by the response. Among 25 frontier models, the best model meets about 88% of weighted criteria, the median about 71%, while clinician reference answers reach about 99%. The authors say the most prominent failure is identifying a crisis but not providing help: in conversations where models passed all identification criteria, about 35% failed at least one actionable resource criterion; the median model scored about 46% on de-escalation and about 8% on disclaimers, while the best model scored about 75% and 85%, respectively.

3. 06:09
   arXiv：危险能力与安全评测Selected

   DistressBench: Evaluating multi-dimensional crisis support in large language models

   Researchers from Scale AI and other institutions released DistressBench, a crisis support evaluation benchmark written by practicing clinicians and crisis counselors, covering 718 English crisis conversations about suicide and self-harm, spanning 23 subcategories and seven support dimensions, with a total of 6,901 weighted criteria. The study evaluated 25 frontier models; the best model met 88.4% of weighted criteria, the median model 71.4%, and clinician reference answers 98.7%. The authors report that the main failure was not harmful outputs but inaction after identifying crises; in conversations where models met all crisis identification criteria, 35.3% did not direct users to human help, and even the strongest model had 15.9%. Models were near saturation in avoiding harmful content (median 94.5%), but de-escalation had a median of only 46.0%, with the best model at 74.6%; multi-turn conversations scored 68.5%, lower than single-turn's 74.2%.

Current attention64·Peak in the comparable period67（Oct 9 22:00）·Change over the comparable past 24 hours +5%