EviSafe - Natural Severity Accuracy: leaderboard
Natural severity accuracy on EviSafe original scenarios with complete parsed victim outputs and GPT-5.4 judge results. Percentages use the per-model N printed in Table 2; instruction-tuned VLM checkpoints are identified by the formal model suite.
Metric: Accuracy (%). Source: arxiv.org. Saturation forecast: Around 2032. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3 VL 4B Instruct | 52.8 | #515 |
| 2 | Qwen 3 VL 235B A22B Instruct | 51.9 | #273 |
| 3 | Qwen 3 VL 8B Instruct | 50.3 | #420 |
| 4 | Gemini 2.5 Pro | 44 | #152 |
| 5 | Gemma 3 12B (IT) | 42.3 | #664 |
| 6 | Claude Opus 4.6 | 41.5 | #63 |
| 7 | Qwen 2.5 VL 7B Instruct | 36.2 | #654 |
| 8 | InternVL3-14B | 35.6 | #505 |
| 9 | InternVL3-8B | 33.4 | #620 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=evisafe-natural-severity-accuracy · How It Works · Data refreshed daily, snapshot 2026-10-20.