Arithmetic Reasoning on GSM8K Aug
83.1AccuracyCoT-SFT
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| CoT-SFTBackbone=Qwen3-4B, Evaluation Protocol=zero-shot2026.06 | 83.1 | 119.8 | |
| CoT-SFTBackbone=Qwen3-VL-4B-Instruct, Evaluation Protocol=zero-shot2026.06 | 81.2 | 127.3 | |
| DLR (Full)Backbone=Qwen3-VL-4B-Instruct, Evaluation Protocol=zero-shot2026.06 | 63.3 | 6.5 | |
| DLR (Full)Backbone=Qwen3-4B, Evaluation Protocol=zero-shot2026.06 | 62.2 | 6.5 | |
| DLR (SFT)Backbone=Qwen3-VL-4B-Instruct, Evaluation Protocol=zero-shot2026.06 | 48 | 6.3 | |
| DLR (SFT)Backbone=Qwen3-4B, Evaluation Protocol=zero-shot2026.06 | 47.4 | 6.4 | |
| CoLaRBackbone=Qwen3-4B, Evaluation Protocol=zero-shot2026.06 | 40 | 39.6 | |
| CoLaRBackbone=Qwen3-VL-4B-Instruct, Evaluation Protocol=zero-shot2026.06 | 39.2 | 38.8 | |
| RoTBackbone=Qwen3-VL-4B-Instruct, Evaluation Protocol=zero-shot2026.06 | 37.8 | 32 | |
| Direct-SFTBackbone=Qwen3-4B, Evaluation Protocol=zero-shot2026.06 | 27.4 | 0 | |
| Direct-SFTBackbone=Qwen3-VL-4B-Instruct, Evaluation Protocol=zero-shot2026.06 | 26.2 | 0 | |
| CoconutBackbone=Qwen3-4B, Evaluation Protocol=zero-shot2026.06 | 16.9 | 6 | |
| CoconutBackbone=Qwen3-VL-4B-Instruct, Evaluation Protocol=zero-shot2026.06 | 16.1 | 6 | |
| iCoTBackbone=Qwen3-4B, Evaluation Protocol=zero-shot2026.06 | 13.5 | 0 | |
| CODIBackbone=Qwen3-4B, Evaluation Protocol=zero-shot2026.06 | 7.3 | 6 | |
| CODIBackbone=Qwen3-VL-4B-Instruct, Evaluation Protocol=zero-shot2026.06 | 7.1 | 6 |