Audio-conditioned reasoning on MMSU
57.63AccCORD
Evaluation results
| Method | Links | ||
|---|---|---|---|
| CORDBackbone=Step-Audio2-mini, Optimization Strategy=CORD2026.01 | 57.63 | 2.84 | |
| SFTBackbone=Step-Audio2-mini, Optimization Strategy=SFT2026.01 | 54.68 | 5.79 | |
| Forward KLBackbone=Step-Audio2-mini, Optimization Strategy=Forward KL2026.01 | 53.12 | 7.35 | |
| Step-Audio2 miniBackbone=Step-Audio2-mini, Optimization Strategy=Base2026.01 | 52.31 | 8.16 | |
| CORDBackbone=Qwen2-Audio-7B-Instruct, Optimization Strategy=CORD2026.01 | 38.06 | 6.4 | |
| Forward KLBackbone=Qwen2-Audio-7B-Instruct, Optimization Strategy=Forward KL2026.01 | 36.77 | 7.69 | |
| SFTBackbone=Qwen2-Audio-7B-Instruct, Optimization Strategy=SFT2026.01 | 36.47 | 7.99 | |
| Qwen2-Audio-7B-InstructBackbone=Qwen2-Audio-7B-Instruct, Optimization Strategy=Base2026.01 | 36.04 | 8.42 |