Grade-school Reasoning on GSM8K Aug
59.7AccuracyCoT-SFT
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| CoT-SFTBackbone=Qwen3-VL-2B-Instruct2026.06 | 59.7 | 131.4 | |
| CoT-SFTBackbone=LLaMA-3.2-1B2026.06 | 54.1 | 123.2 | |
| DLR (Full)Backbone=Qwen3-VL-2B-Instruct2026.06 | 53.4 | 6 | |
| DLR (Full)Backbone=LLaMA-3.2-1B2026.06 | 43.4 | 6.4 | |
| ReGuLaRBackbone=LLaMA-3.2-1B2026.06 | 34.9 | 3.7 | |
| CoLaRBackbone=Qwen3-VL-2B-Instruct2026.06 | 28.3 | 5.7 | |
| CoLaRBackbone=LLaMA-3.2-1B2026.06 | 26.6 | 5.6 | |
| RoTBackbone=Qwen3-VL-2B-Instruct2026.06 | 23.3 | 32 | |
| CoconutBackbone=LLaMA-3.2-1B2026.06 | 20.5 | 6 | |
| iCoTBackbone=LLaMA-3.2-1B2026.06 | 19.8 | 0 | |
| Direct-SFTBackbone=Qwen3-VL-2B-Instruct2026.06 | 15.6 | 0 | |
| Direct-SFTBackbone=LLaMA-3.2-1B2026.06 | 13.9 | 0 | |
| CoconutBackbone=Qwen3-VL-2B-Instruct2026.06 | 13.4 | 6 | |
| CODIBackbone=LLaMA-3.2-1B2026.06 | 13.3 | 6 | |
| CODIBackbone=Qwen3-VL-2B-Instruct2026.06 | 9.3 | 6 |