aiXamine - Over-Refusal: leaderboard
aiXamine Over-Refusal service score: answering benign queries that resemble unsafe ones instead of refusing excessively (aiXamine, arXiv 2608.20554v1, Appendix Table 7). Normalized 0-100 score aggregated from the service's tests; higher is better. 42 models evaluated under identical black-box conditions via OpenRouter or containerized vLLM.
Metric: Service score (0-100; normalized pass rate). Source: arxiv.org. Saturation forecast: Estimated already saturated. 42 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemma 3 27B | 94.6 | #606 |
| 2 | GPT-4o | 94.4 | #357 |
| 3 | Gemini 3 Flash (Preview) | 94.4 | #90 |
| 4 | Grok 3 Mini | 94.4 | #289 |
| 5 | Llama 4 Scout | 94 | #648 |
| 6 | Llama 3.3 70B | 94 | #578 |
| 7 | Llama 3.2 3B | 93.8 | #1376 |
| 8 | Gemma 3 12B | 93.1 | #674 |
| 9 | DeepSeek R1 0528 | 91.5 | #232 |
| 10 | Gemini 2.5 Flash | 90.4 | #255 |
| 11 | Llama 3.1 8B | 90.3 | #1130 |
| 12 | Gemma 3 4B | 90 | #1088 |
| 13 | DeepSeek R1 Distill Llama 8B | 89.2 | #1295 |
| 14 | Grok 4 Fast | 88.3 | #266 |
| 15 | Grok 4.1 Fast | 88.2 | #227 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=aixamine-over-refusal · How It Works · Data refreshed daily, snapshot 2026-11-05.