Multiple-choice Video Harmfulness Understanding on HarmVideoBench (test)
98.9Observable Evidence AccuracyHuman
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| HumanSize=-2026.06 | 98.9 | 97.5 | 96.2 | 97.5 | |
| BCRSize=8B, Base Model=Qwen3-VL-8B, Mode=Task-adapted2026.06 | 96.7 | 86.4 | 70.1 | 84.4 | |
| Gemini-3.1-ProSize=-, Source=closed-source2026.06 | 96 | 84.6 | 65.8 | 82.1 | |
| Gemini-3.1-FlashSize=-, Source=closed-source2026.06 | 95.7 | 83.7 | 64.4 | 81.3 | |
| GPT-5.5Size=-, Source=closed-source2026.06 | 95.4 | 83.4 | 63.9 | 80.9 | |
| Qwen3-VL-235B-A22B-ThinkingSize=235B-A22B, Source=Open-Source, Mode=thinking2026.06 | 95.1 | 82.6 | 62.5 | 80.1 | |
| Qwen3-VL-235B-A22B-InstructSize=235B-A22B, Source=Open-Source, Mode=non-thinking2026.06 | 94.8 | 82.3 | 62 | 79.7 | |
| Claude-Opus-4.7Size=-, Source=closed-source2026.06 | 94.7 | 82.9 | 63.2 | 80.3 | |
| Qwen3-VL-32B-ThinkingSize=32B, Source=Open-Source, Mode=thinking2026.06 | 94.7 | 82.1 | 62.1 | 79.6 | |
| Claude-Sonnet-4.6Size=-, Source=closed-source2026.06 | 94.4 | 82.5 | 62.7 | 79.9 | |
| Qwen3-VL-32B-InstructSize=32B, Source=Open-Source, Mode=non-thinking2026.06 | 94.3 | 81.6 | 61.1 | 79 | |
| Qwen3-VL-30B-A3B-ThinkingSize=30B-A3B, Source=Open-Source, Mode=thinking2026.06 | 94 | 81.4 | 61.2 | 78.9 | |
| HarmVideo-SFTSize=8B, Base Model=Qwen3-VL-8B, Mode=Task-adapted2026.06 | 93.5 | 83 | 64.8 | 80.5 | |
| Phi-4-reasoning-vision-15BSize=15B, Source=Open-Source, Mode=non-thinking2026.06 | 93 | 80.9 | 61.2 | 78.4 | |
| Qwen3-VL-30B-A3B-InstructSize=30B-A3B, Source=Open-Source, Mode=non-thinking2026.06 | 91.5 | 78.9 | 58.5 | 76.3 | |
| Qwen3-VL-8B-ThinkingSize=8B, Source=Open-Source, Mode=thinking2026.06 | 91.2 | 79.4 | 59.5 | 76.7 | |
| Kimi-K2.5Size=170B, Source=Open-Source, Mode=non-thinking2026.06 | 85 | 73.9 | 54.8 | 71.2 | |
| MiniCPM-V-4.6Size=1.3B, Source=Open-Source, Mode=non-thinking2026.06 | 83.9 | 72.4 | 53.3 | 69.9 | |
| Lance-3BSize=3B, Source=Open-Source, Mode=non-thinking2026.06 | 82.1 | 71 | 52.3 | 68.5 | |
| LFM2.5-VL-1.6BSize=1.6B, Source=Open-Source, Mode=non-thinking2026.06 | 79.8 | 69.3 | 51.4 | 66.8 | |
| GLM-4.5VSize=106B-A12B, Source=Open-Source, Mode=non-thinking2026.06 | 77.7 | 67.8 | 50.9 | 65.5 | |
| Qwen3-VL-8B-InstructSize=8B, Source=Open-Source, Mode=non-thinking2026.06 | 73.3 | 63.7 | 48.1 | 61.7 | |
| RandomSize=-2026.06 | 24.9 | 24.9 | 25.1 | 25 |