Audio-conditioned reasoning on OBQA
77.74AccuracyCORD
Evaluation results
| Method | Links | ||
|---|---|---|---|
| CORDBackbone=Step-Audio2-mini, Optimization Strategy=CORD2026.01 | 77.74 | 2.26 | |
| SFTBackbone=Step-Audio2-mini, Optimization Strategy=SFT2026.01 | 74.72 | 5.28 | |
| Step-Audio2 miniBackbone=Step-Audio2-mini, Optimization Strategy=Base2026.01 | 72.3 | 7.7 | |
| Forward KLBackbone=Step-Audio2-mini, Optimization Strategy=Forward KL2026.01 | 72.08 | 7.92 | |
| CORDBackbone=Qwen2-Audio-7B-Instruct, Optimization Strategy=CORD2026.01 | 52.77 | 16.02 | |
| Qwen2-Audio-7B-InstructBackbone=Qwen2-Audio-7B-Instruct, Optimization Strategy=Base2026.01 | 51.2 | 17.59 | |
| SFTBackbone=Qwen2-Audio-7B-Instruct, Optimization Strategy=SFT2026.01 | 49.49 | 19.3 | |
| Forward KLBackbone=Qwen2-Audio-7B-Instruct, Optimization Strategy=Forward KL2026.01 | 48.21 | 20.58 |