Multimodal Reasoning on MathVision (Acc, AUROC, ECE)
28.2AccuracySC
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| SCBackbone=Qwen2.5VL-7B-Instruct, Averaged Over Seeds=false2026.07 | 28.2 | 70 | 19.3 | |
| C3RL (Ours)Backbone=Qwen2.5VL-7B-Instruct, Averaged Over Seeds=true2026.07 | 27.9 | 62.7 | 29 | |
| C3RL w/o RefBackbone=Qwen2.5VL-7B-Instruct, Averaged Over Seeds=true2026.07 | 27.3 | 60.7 | 38.2 | |
| RLCRBackbone=Qwen2.5VL-7B-Instruct, Averaged Over Seeds=true2026.07 | 25.4 | 62.6 | 25.3 | |
| RLVRBackbone=Qwen2.5VL-7B-Instruct, Averaged Over Seeds=true2026.07 | 24.6 | 51.2 | 73.8 | |
| BaseBackbone=Qwen2.5VL-7B-Instruct, Averaged Over Seeds=true2026.07 | 24.2 | 50.8 | 68 | |
| SaySelfBackbone=Qwen2.5VL-7B-Instruct, Averaged Over Seeds=true2026.07 | 18.4 | 69.4 | 13.3 | |
| SFT+RefBackbone=Qwen2.5VL-7B-Instruct, Averaged Over Seeds=true2026.07 | 18.1 | 61.3 | 8.3 |