TrustMH-Bench - Crisis Escalation: leaderboard
The model answers user messages showing one of six mental-health crises (or none), and a GPT-4.1 judge rates 1-5 whether its reply escalates and refers to help at the level an expert protocol requires; the number is the mean rating. Unlike the TrustMH-Bench jailbreak board, which scores refusing harmful requests, it scores the response to a person in crisis.
Metric: Mean escalation rating (1-5, higher is better). Source: arxiv.org. 12 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.1 | 4.54 | #229 |
| 2 | DeepSeek V3.2 | 4.22 | #406 |
| 3 | Claude Sonnet 4.5 | 4.11 | #250 |
| 4 | Qwen 3 235B A22B 2507 Instruct | 4.08 | #619 |
| 5 | Gemini 2.5 Flash | 4.08 | #453 |
| 6 | GPT-4o Mini | 3.86 | #1623 |
Interactive version: theaggregate.ai/benchmark?slug=trustmh-bench-crisis-escalation · How It Works · Data refreshed daily, snapshot 2026-11-13.