Mathematical Reasoning on GSM (Acc, # Tokens)
61AccuracyL-ImgCoT
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| L-ImgCoTBackbone=LLama3.2-3B-Instruction2026.01 | 61 | 63.2 | |
| Full-CoTBackbone=LLama3.2-3B-Instruction2026.01 | 60.5 | 105.6 | |
| ImgCoTBackbone=LLama3.2-3B-Instruction2026.01 | 56.8 | 8 | |
| ImgCoT (w/o layout)Backbone=LLama3.2-3B-Instruction2026.01 | 52.7 | 8 | |
| ImgCoT (w/ textual tokens)Backbone=LLama3.2-3B-Instruction2026.01 | 49.3 | 8 | |
| ICoTBackbone=LLama3.2-3B-Instruction2026.01 | 46.7 | 0 | |
| L-ImgCoTBackbone=Qwen2.5-1.5B-Instruction2026.01 | 45.2 | 71.3 | |
| Full-CoTBackbone=Qwen2.5-1.5B-Instruction2026.01 | 44.1 | 126.6 | |
| ImgCoTBackbone=Qwen2.5-1.5B-Instruction2026.01 | 38.7 | 8 | |
| ImgCoT (w/o layout)Backbone=Qwen2.5-1.5B-Instruction2026.01 | 38.4 | 8 | |
| ImgCoT (w/ textual tokens)Backbone=Qwen2.5-1.5B-Instruction2026.01 | 37.5 | 8 | |
| CoLaRBackbone=LLama3.2-3B-Instruction2026.01 | 27.4 | 20.1 | |
| CODIBackbone=LLama3.2-3B-Instruction2026.01 | 23.3 | 6 | |
| ICoTBackbone=Qwen2.5-1.5B-Instruction2026.01 | 22.3 | 0 | |
| CoconutBackbone=LLama3.2-3B-Instruction2026.01 | 19.5 | 6 | |
| L-ImgCoTBackbone=Qwen2.5-0.5B-Instruction2026.01 | 17.5 | 64.7 | |
| Full-CoTBackbone=Qwen2.5-0.5B-Instruction2026.01 | 16.9 | 116.3 | |
| ImgCoTBackbone=Qwen2.5-0.5B-Instruction2026.01 | 9.2 | 8 | |
| ImgCoT (w/o layout)Backbone=Qwen2.5-0.5B-Instruction2026.01 | 9.2 | 8 | |
| ImgCoT (w/ textual tokens)Backbone=Qwen2.5-0.5B-Instruction2026.01 | 8.3 | 8 | |
| CoLaRBackbone=Qwen2.5-1.5B-Instruction2026.01 | 6.9 | 22.9 | |
| CoLaRBackbone=Qwen2.5-0.5B-Instruction2026.01 | 5.8 | 19.1 | |
| CoconutBackbone=Qwen2.5-1.5B-Instruction2026.01 | 5.4 | 6 | |
| ICoTBackbone=Qwen2.5-0.5B-Instruction2026.01 | 3.9 | 0 | |
| CoconutBackbone=Qwen2.5-0.5B-Instruction2026.01 | 3.7 | 6 | |
| CODIBackbone=Qwen2.5-1.5B-Instruction2026.01 | 3.6 | 6 | |
| CODIBackbone=Qwen2.5-0.5B-Instruction2026.01 | 0.5 | 6 |