AVI-Bench - Visual-Referenced Audio Hallucination: leaderboard
An AVI-Bench task: visual-referenced audio hallucination (250 items): accuracy of the yes, no or not sure verdict on whether a visually described sound is present. AVI-Bench (2026) evaluates omni-modal models zero-shot on audio-visual tasks arranged by cognitive stage; scores in percent, higher is better.
Metric: Task score (%) on visual-referenced audio hallucination (250 items): accuracy of the yes, no or not sure verdict on whether a visually described sound is present, on AVI-Bench (5,864 audio-visual items in 14 tasks over four cognitive stages), zero-shot with the same prompts for every omni-modal model; task scores normalized to percentages; higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 26 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Pro (Preview 05-06) | 80.8 | #275 |
| 2 | Qwen2.5-Omni-7B | 77.6 | #1789 |
| 3 | Gemini 2.0 Flash | 75.2 | #704 |
| 4 | Gemini 2.5 Flash (Preview 04-17) | 70.8 | #450 |
| 5 | Gemini 1.5 Flash | 66.4 | #1668 |
| 6 | Gemini 1.5 Pro | 62.8 | #1049 |
| 7 | Phi-4 Multimodal Instruct | 60.4 | #3882 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=avi-bench-visual-referenced-audio-hallucination · How It Works · Data refreshed daily, snapshot 2026-11-13.