Arithmetic Reasoning on MultiArith (Accuracy and #L)
98.5AccuracyCoT-SFT
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| CoT-SFTBackbone=Qwen3-4B2026.06 | 98.5 | 57.6 | |
| CoT-SFTBackbone=Qwen3-VL-4B-Instruct2026.06 | 98.3 | 59.1 | |
| DLR (Full)Backbone=Qwen3-VL-4B-Instruct2026.06 | 98.3 | 4.3 | |
| DLR (Full)Backbone=Qwen3-4B2026.06 | 97.8 | 4.3 | |
| DLR (SFT)Backbone=Qwen3-VL-4B-Instruct2026.06 | 97.8 | 4.2 | |
| RoTBackbone=Qwen3-VL-4B-Instruct2026.06 | 97.2 | 32 | |
| DLR (SFT)Backbone=Qwen3-4B2026.06 | 96.1 | 4.2 | |
| Direct-SFTBackbone=Qwen3-4B2026.06 | 86.3 | 0 | |
| Direct-SFTBackbone=Qwen3-VL-4B-Instruct2026.06 | 85.6 | 0 | |
| CoLaRBackbone=Qwen3-4B2026.06 | 82.2 | 21.1 | |
| CoLaRBackbone=Qwen3-VL-4B-Instruct2026.06 | 81 | 22.1 | |
| CoconutBackbone=Qwen3-4B2026.06 | 60.3 | 6 | |
| CoconutBackbone=Qwen3-VL-4B-Instruct2026.06 | 58.9 | 6 | |
| iCoTBackbone=Qwen3-4B2026.06 | 49.2 | 0 | |
| CODIBackbone=Qwen3-4B2026.06 | 18.3 | 6 | |
| CODIBackbone=Qwen3-VL-4B-Instruct2026.06 | 17.8 | 6 |