Researchers have developed DistressBench, a new evaluation dataset for assessing how well large language models (LLMs) can provide support in crisis situations, particularly those involving suicide and self-harm. Unlike existing benchmarks that focus on refusal, DistressBench uses a rubric of weighted criteria across seven dimensions of crisis support, including de-escalation and routing users to human help. The evaluation found that even the best-performing frontier models only satisfied 88.4% of the weighted criteria, with a significant portion failing to route users to appropriate resources despite recognizing the crisis. The study emphasizes the need to evaluate LLMs on the quality of care they deliver rather than just their ability to withhold responses. AI
IMPACT Establishes a new standard for evaluating LLM safety in sensitive crisis scenarios, pushing for more nuanced support capabilities beyond simple refusal.
RANK_REASON The item describes a new academic paper introducing a novel benchmark for evaluating LLMs. [lever_c_demoted from research: ic=1 ai=1.0]
- alphaXiv
- arXiv
- CatalyzeX
- DagsHub
- DistressBench
- Gotit.pub
- Hugging Face
- large-language models
- ScienceCast
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
