Math Reasoning on MultiArith (Accuracy and Reasoning Length)
99.3AccuracyCoT-SFT
Evaluation results
| Method | Links | ||
|---|---|---|---|
| CoT-SFTBackbone=LLaMA-3.2-1B2026.06 | 99.3 | 69.4 | |
| SFT-CoTModel=Qwen3-VL-4B-Instruct2026.01 | 98.3 | 59.1 | |
| RoT (Ours)Model=Qwen3-VL-4B-Instruct2026.01 | 97.2 | 32 | |
| SFT-CoTModel=Qwen3-VL-2B-Instruct2026.01 | 95 | 68 | |
| CoT-SFTBackbone=Qwen3-VL-2B-Instruct2026.06 | 95 | 68 | |
| DLR (Full)Backbone=Qwen3-VL-2B-Instruct2026.06 | 94.4 | 4.2 | |
| DLR (Full)Backbone=LLaMA-3.2-1B2026.06 | 93.7 | 4.1 | |
| ReGuLaRBackbone=LLaMA-3.2-1B-Instruct2026.01 | 89.2 | 2.28 | |
| ReGuLaRBackbone=LLaMA-3.2-1B2026.06 | 89.2 | 2.3 | |
| CoLaRBackbone=LLaMA-3.2-1B-Instruct, Compression Rate=52026.01 | 87 | 3.23 | |
| CoLaRBackbone=LLaMA-3.2-1B2026.06 | 87 | 3.2 | |
| SFT-CoTModel=LLaVa-V1.6-Mistral-7B2026.01 | 86.1 | 82.3 | |
| SFT-w/o CoTModel=Qwen3-VL-4B-Instruct2026.01 | 85.6 | 0 | |
| CoLaRBackbone=Qwen3-VL-2B-Instruct2026.06 | 79.4 | 3.4 | |
| RoT (Ours)Model=LLaVa-V1.6-Mistral-7B2026.01 | 68.3 | 32 | |
| RoT (Ours)Model=Qwen3-VL-2B-Instruct2026.01 | 62.2 | 32 | |
| RoTBackbone=Qwen3-VL-2B-Instruct2026.06 | 62.2 | 32 | |
| CoconutBackbone=Qwen3-VL-2B-Instruct2026.06 | 52.9 | 6 | |
| SFT-w/o CoTModel=LLaVa-V1.6-Mistral-7B2026.01 | 52.8 | 0 | |
| SFT-w/o CoTModel=Qwen3-VL-2B-Instruct2026.01 | 41.7 | 0 | |
| Direct-SFTBackbone=Qwen3-VL-2B-Instruct2026.06 | 41.7 | 0 | |
| CoconutBackbone=LLaMA-3.2-1B-Instruct2026.01 | 41.4 | 6 | |
| CoconutBackbone=LLaMA-3.2-1B2026.06 | 41.4 | 6 | |
| iCoTBackbone=LLaMA-3.2-1B-Instruct2026.01 | 38.2 | 0 | |
| iCoTBackbone=LLaMA-3.2-1B2026.06 | 38.2 | 0 | |
| Direct-SFTBackbone=LLaMA-3.2-1B2026.06 | 32.2 | 0 | |
| CODIBackbone=LLaMA-3.2-1B-Instruct2026.01 | 19.2 | 6 | |
| CODIBackbone=LLaMA-3.2-1B2026.06 | 19.2 | 6 | |
| CODIBackbone=Qwen3-VL-2B-Instruct2026.06 | 17.5 | 6 |