LetTokenBurnDistressBench: Evaluating Multi-Dimensional Crisis Support in Large Language Models
Evaluates LLM crisis support via 718 clinician-authored SSH conversations scored against weighted atomic criteria across seven dimensions.
$ npm i -g lettokenburn-cli$ ltb bench info distressbench
External harness. No Harbor adapter has been generated for this benchmark yet. `ltb bench info` prints the known links.
Runs happen on your machine with your own subscription; only the score summary is pushed. CLI docs
| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| Muse Spark 1.2 | 0.9 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
| Muse Spark 1.3 | 0.8 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
| Muse Glimmer 30B | 0.8 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
| Inkling | 0.8 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
| Qwen3.8 Max | 0.8 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
| GPT-5.6-terra | 0.8 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
| Kimi K2P6 | 0.8 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
| GPT-5.6 Sol | 0.7 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
| Kimi K3 | 0.7 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
| Qwen3.7 Plus | 0.7 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
| GPT-5.6 Luna | 0.7 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
| Claude Fable 5.1 | 0.7 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
| Claude Opus 4.8 | 0.7 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
| GLM-5 | 0.7 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.
DistressBench comprises 718 English conversations covering 23 suicide and self-harm subcategories, authored by clinicians with specific rubrics and reference responses. It includes single-turn and multi-turn tasks designed to simulate user distress scenarios.
Scoring uses a frozen three-judge LLM panel calibrated on clinician labels to compute a weighted average of satisfied atomic criteria. Results show models struggle with action-oriented dimensions like de-escalation and resource routing compared to recognition.