Arithmetic Reasoning on SVAMP (Acc. & #L)
84.3AccuracyCoT-SFT
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| CoT-SFTBackbone=Qwen3-VL-4B-Instruct2026.06 | 84.3 | 55.9 | |
| CoT-SFTBackbone=Qwen3-4B2026.06 | 82.9 | 53.7 | |
| RoTBackbone=Qwen3-VL-4B-Instruct2026.06 | 72.7 | 32 | |
| DLR (Full)Backbone=Qwen3-VL-4B-Instruct2026.06 | 72.7 | 3.8 | |
| DLR (Full)Backbone=Qwen3-4B2026.06 | 72.3 | 3.8 | |
| DLR (SFT)Backbone=Qwen3-4B2026.06 | 70.3 | 3.6 | |
| Direct-SFTBackbone=Qwen3-VL-4B-Instruct2026.06 | 70 | 0 | |
| Direct-SFTBackbone=Qwen3-4B2026.06 | 69.4 | 0 | |
| DLR (SFT)Backbone=Qwen3-VL-4B-Instruct2026.06 | 68.7 | 3.6 | |
| CoLaRBackbone=Qwen3-4B2026.06 | 57.7 | 19.2 | |
| CoLaRBackbone=Qwen3-VL-4B-Instruct2026.06 | 52 | 18.4 | |
| CoconutBackbone=Qwen3-4B2026.06 | 43.6 | 6 | |
| CoconutBackbone=Qwen3-VL-4B-Instruct2026.06 | 42 | 6 | |
| iCoTBackbone=Qwen3-4B2026.06 | 36.9 | 0 | |
| CODIBackbone=Qwen3-4B2026.06 | 11 | 6 | |
| CODIBackbone=Qwen3-VL-4B-Instruct2026.06 | 10.7 | 6 |