AOR-Bench: leaderboard
Audio over-refusal benchmark (AOR-Bench, Yang et al., 2026): 500 pseudo-harmful audio queries whose speech sounds harmful on its own but whose background audio (from 50 scenes in six scenario categories) makes the request benign. The score is the over-refusal rate, the percentage of these queries an audio language model refuses, judged by an LLM; lower is better.
Metric: Over-refusal rate (%; refusals of 500 pseudo-harmful audio queries, pooled over six scenario categories). Source: arxiv.org. Saturation forecast: Around February 2027. 12 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.0 Flash Lite | 38.2 | #1032 |
| 2 | Gemini 3 Flash (Preview) | 43.8 | #141 |
| 3 | Gemini 2.5 Flash Lite | 66.2 | #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 · How It Works · Data refreshed daily, snapshot 2026-11-12.