SVHalluc - Cross-Modal Semantic Binding: leaderboard
Speech-vision hallucination (SVHalluc, Zhang et al., CVPR 2026): 534 yes/no questions on whether an event composed from spoken and visual content actually appears in the video, for audio-visual LLMs on cooking videos with narration. Option-balanced; accuracy in percent, higher is better.
Metric: Accuracy (%; 534 option-balanced video questions, whether an event composed from speech and vision is visible (yes/no)). Source: arxiv.org. Saturation forecast: Around February 2027. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Pro | 78.56 | #262 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=svhalluc-cross-modal-semantic-binding · How It Works · Data refreshed daily, snapshot 2026-11-12.