Olmo 3 Safety Evaluation - XSTest: leaderboard
The model gets 200 unsafe prompts and 250 safe prompts that share their wording (homonyms, figurative language, fiction); the score is the share handled correctly, refusing the unsafe and answering the safe. Unlike HELM's XSTest board, this re-run covers 2025 open models with a WildGuard refusal judge.
Metric: Accuracy: refuse 200 unsafe, comply with 250 safe look-alike prompts (%, higher is better). Source: arxiv.org. 36 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Olmo-3.1-32B-Instruct-DPO | 95.1 | #3174 |
| 2 | Qwen 3 VL 8B (Thinking) | 94.2 | #887 |
| 3 | Qwen 2.5 32B Instruct | 94 | #1215 |
| 4 | OLMo-2-1124-7B-Instruct | 93.9 | #7343 |
| 5 | OLMo 3 32B (Thinking) | 93.9 | #1884 (OLMo 3 32B) |
| 6 | Qwen 2.5 7B Instruct | 93.8 | #2564 |
| 7 | Qwen 3 VL 32B Instruct | 93.7 | #585 |
| 8 | Qwen 3 VL 32B (Thinking) | 93.4 | #559 |
| 9 | Qwen 3 VL 8B Instruct | 93.2 | #981 |
| 10 | OLMo 3 7B Instruct | 93.2 | #6860 |
| 11 | OLMo 3.1 32B Instruct | 93.1 | #2384 |
| 12 | Olmo-3.1-32B-Instruct-SFT | 93 | #3407 |
| 13 | OLMo-2-0325-32B-Instruct | 92.7 | #2570 |
| 14 | NVIDIA Nemotron Nano 9B V2 | 92.5 | #2270 |
| 15 | Gemma 3 27B (IT) | 92.1 | #1289 |
Interactive version: theaggregate.ai/benchmark?slug=olmo-3-safety-evaluation-xstest · How It Works · Data refreshed daily, snapshot 2026-11-13.