Multimodal Hallucination - AVH Visual SelfCheckGPT Resistance: leaderboard
The "AVH Visual SelfCheckGPT Resistance" subtask of Multimodal Hallucination. A task measuring hallucination in multimodal models — how often a vision-language model asserts objects, attributes or facts not supported by the image.
Metric: Resistance (100 - SelfCheckGPT). Source: huggingface.co. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini-v1.5-pro (audio-only & audio-visual) | 87.23 | |
| 2 | Gemini-v1.5-pro (visual-only) | 80.13 | |
| 3 | Chat-Univi | 70.6 | |
| 4 | Video-LLaVA | 69.41 | |
| 5 | Llama-VID | 61.39 | |
| 6 | Video-Llama (visual-only) | 58.86 | |
| 7 | Video-Llama (audio-only & audio-visual) | 52.27 | |
| 8 | Valley | 47.57 | |
| 9 | FAVOR (audio-only & audio-visual) | 43.58 | |
| 10 | FAVOR (visual-only) | 39.33 |
Interactive version: theaggregate.ai/benchmark?slug=multimodal-hallucination-avh-visual-selfcheckgpt-resistance · How It Works · Data refreshed daily, snapshot 2026-10-20.