RedVox - Italian: leaderboard
The Italian requests of RedVox (Savoldi et al., 2026): harmful and stereotyped requests recorded by native speakers, spoken or written with a silent or noisy audio clip attached. Share of responses a GPT-5.5 judge labels unsafe, printed as a whole percent truncated toward zero; lower is better.
Metric: Unsafe response rate (whole percent, truncated as printed; share of responses a GPT-5.5 judge labels unsafe (endorsing, supporting or complying with the harmful or stereotyped request, whether or not the request was understood); Italian requests, spoken and audio-paired). Source: arxiv.org. Saturation forecast: Estimated already saturated. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Flash Lite | 2 | #363 |
| 2 | Qwen3 Omni 30B A3B Instruct | 2 | #902 |
| 3 | Gemini 3.1 Pro (Preview) | 3 | #91 |
| 4 | Phi-4 Multimodal Instruct | 17 | #3914 |
| 5 | Voxtral-Small-24B-2507 | 26 | #1291 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=redvox-italian · How It Works · Data refreshed daily, snapshot 2026-11-12.