Paralinguistic Reasoning on SAKURA
92.2Animal Accuracy (Single-hop)Qwen2.5-Omni
Evaluation Results
| Method | Links | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| Qwen2.5-OmniParam=3B, Audio (hrs)=∼1.7M–3.3M est.2026.07 | 92.2 | 70.6 | 64.8 | 22.6 | 27.4 | 15.4 | 80 | 29 | 66.1 | 34.4 | |
| QWEN2-AUDIOParam=7B, Audio (hrs)=∼520k2026.07 | 90 | 60.6 | 89 | 43.4 | 62.2 | 40.2 | 83.4 | 46.4 | 81.2 | 47.7 | |
| Audio Flamingo-3Param=8B, Audio (hrs)=∼100k2026.07 | 88 | 61.6 | 66.4 | 49.4 | 70.8 | 41.2 | 93.8 | 43.6 | 79.8 | 49 | |
| ORCAParam=4B, Audio (hrs)=7,0002026.07 | 76 | 76.2 | 87.6 | 85.8 | 63.8 | 57.2 | 94.4 | 81.6 | 80.5 | 75.2 | |
| DESTA2.5-AUDIO (8B)Param=8B, Audio (hrs)=7,0002026.07 | 58.8 | 58 | 89.2 | 84 | 62.2 | 57.4 | 96.4 | 80 | 76.7 | 69.9 | |
| DESTA2.5-AUDIO 4B BaselineParam=4B, Audio (hrs)=7,0002026.07 | 30.8 | 34.2 | 44.2 | 38 | 56 | 50.8 | 93 | 72.2 | 56 | 48.8 |