Audio-conditioned Reasoning on GSM8K
47.56AccuracyCORD
Evaluation results
| Method | Links | ||
|---|---|---|---|
| CORDBackbone=Step-Audio2-mini, Optimization Strategy=CORD2026.01 | 47.56 | 12.91 | |
| Forward KLBackbone=Step-Audio2-mini, Optimization Strategy=Forward KL2026.01 | 46.57 | 13.9 | |
| Step-Audio2 miniBackbone=Step-Audio2-mini, Optimization Strategy=Base2026.01 | 43.75 | 16.72 | |
| CORDBackbone=Qwen2-Audio-7B-Instruct, Optimization Strategy=CORD2026.01 | 36.2 | 4.27 | |
| Forward KLBackbone=Qwen2-Audio-7B-Instruct, Optimization Strategy=Forward KL2026.01 | 36.05 | 4.42 | |
| SFTBackbone=Qwen2-Audio-7B-Instruct, Optimization Strategy=SFT2026.01 | 33.68 | 6.79 | |
| SFTBackbone=Step-Audio2-mini, Optimization Strategy=SFT2026.01 | 27.89 | 32.58 | |
| Qwen2-Audio-7B-InstructBackbone=Qwen2-Audio-7B-Instruct, Optimization Strategy=Base2026.01 | 20.73 | 19.74 |