Multimodal Reasoning on MathVista (Acc, AUROC, ECE)
62.2AccuracyC3RL w/o Ref
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| C3RL w/o RefBackbone=Qwen2.5VL-7B-Instruct, Averaged Over Seeds=true2026.07 | 62.2 | 66.9 | 13.9 | |
| C3RL (Ours)Backbone=Qwen2.5VL-7B-Instruct, Averaged Over Seeds=true2026.07 | 62.2 | 68.2 | 7.4 | |
| C3RL w/o RefBackbone=Qwen2.5VL-7B-Instruct, Evaluation Protocol=Averaged over 3 seeds (0, 42, 2025)2026.07 | 62.2 | 66.9 | 13.9 | |
| C3RLBackbone=Qwen2.5VL-7B-Instruct, Evaluation Protocol=Averaged over 3 seeds (0, 42, 2025)2026.07 | 62.2 | 68.2 | 7.4 | |
| RLVRBackbone=Qwen2.5VL-7B-Instruct, Averaged Over Seeds=true2026.07 | 61.4 | 59.5 | 30.4 | |
| RLVRBackbone=Qwen2.5VL-7B-Instruct, Evaluation Protocol=Averaged over 3 seeds (0, 42, 2025)2026.07 | 61.4 | 59.5 | 30.4 | |
| RLCRBackbone=Qwen2.5VL-7B-Instruct, Averaged Over Seeds=true2026.07 | 60 | 69 | 11.7 | |
| RLCRBackbone=Qwen2.5VL-7B-Instruct, Evaluation Protocol=Averaged over 3 seeds (0, 42, 2025)2026.07 | 60 | 69 | 11.7 | |
| SCBackbone=Qwen2.5VL-7B-Instruct, Averaged Over Seeds=false2026.07 | 59.1 | 81.8 | 7.7 | |
| SCBackbone=Qwen2.5VL-7B-Instruct, Evaluation Protocol=Single test2026.07 | 59.1 | 81.8 | 7.7 | |
| BaseBackbone=Qwen2.5VL-7B-Instruct, Averaged Over Seeds=true2026.07 | 54.3 | 62.3 | 28.4 | |
| BaseBackbone=Qwen2.5VL-7B-Instruct, Evaluation Protocol=Averaged over 3 seeds (0, 42, 2025)2026.07 | 54.3 | 62.3 | 28.4 | |
| SFT+RefBackbone=Qwen2.5VL-7B-Instruct, Averaged Over Seeds=true2026.07 | 45.2 | 70.2 | 9.1 | |
| SFT+RefBackbone=Qwen2.5VL-7B-Instruct, Evaluation Protocol=Averaged over 3 seeds (0, 42, 2025)2026.07 | 45.2 | 70.2 | 9.1 | |
| SaySelfBackbone=Qwen2.5VL-7B-Instruct, Averaged Over Seeds=true2026.07 | 43.1 | 82.2 | 9.4 | |
| SaySelfBackbone=Qwen2.5VL-7B-Instruct, Evaluation Protocol=Averaged over 3 seeds (0, 42, 2025)2026.07 | 43.1 | 82.2 | 9.4 |