Audio-Visual Hallucination Evaluation on AVHBench
81.29AVGSFFL
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| SFFLBackbone=Qwen3-Omni-30B-A3B-Instruct2026.05 | 81.29 | 80.79 | 85.12 | 79.58 | |
| Qwen3-Omni-ThinkingBackbone=Qwen3-Omni-Thinking2026.05 | 75.95 | 77.12 | 82.31 | 70.68 | |
| PEM-AVQA-14K DATA (GRPO)Backbone=Qwen3-Omni-30B-A3B-Instruct, Protocol=GRPO Training2026.05 | 75.84 | 75.2 | 81.69 | 73.08 | |
| Zero-Shot InferenceBackbone=Qwen3-Omni-30B-A3B-Instruct2026.05 | 73.12 | 74.28 | 81.95 | 66.36 | |
| Gemini-3-FlashBackbone=Gemini-3-Flash2026.05 | 72.27 | 72.14 | 71.65 | 72.81 | |
| VideoLLaMA2-AVCDBackbone=VideoLLaMA2-AVCD2026.05 | 72.15 | — | — | — | |
| SFFLBackbone=Qwen2.5-Omni-7B2026.05 | 64.79 | 62.27 | 78.61 | 59.49 | |
| VideoLLaMA2.1-7B-AVBackbone=VideoLLaMA2.1-7B-AV2026.05 | 63.47 | 73.93 | 61.71 | 51.76 | |
| Zero-Shot InferenceBackbone=Qwen2.5-Omni-7B2026.05 | 63.29 | 61.41 | 70.02 | 61.51 | |
| Video-SALMONN-2+(7B)Backbone=Video-SALMONN-2+(7B)2026.05 | 59.94 | 56.27 | 84.26 | 49.68 |