AOR-Bench - MB-Score: leaderboard
Refusal balance on AOR-Bench (Yang et al., 2026): the harmonic mean of the true-refusal rate on 500 harmful speech-only queries and the compliance rate on their 500 pseudo-harmful counterparts, whose background audio makes the request benign, pooled over six scenario categories. In percent; higher is better.
Metric: MB-Score (%; harmonic mean of the true-refusal rate on 500 harmful speech-only queries and 100 minus the over-refusal rate, pooled over six scenario categories). Source: arxiv.org. Saturation forecast: Around August 2027. 12 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Flash (Preview) | 55.9 | #141 |
| 2 | Gemini 2.0 Flash Lite | 54.16 | #1032 |
| 3 | Gemini 2.5 Flash Lite | 47.97 | #958 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=aor-bench-mb-score · How It Works · Data refreshed daily, snapshot 2026-11-12.