Multimodal Mathematical Reasoning on WeMath
36.33WeMath-S ScoreGAP (LH+PCA+DA)
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| GAP (LH+PCA+DA)Latent Head (LH)=true, Difficulty-Aware supervision (DA)=true, PCA alignment (PCA)=true2026.05 | 36.33 | 54.57 | 53.97 | |
| Qwen2.5-VL 7BParameters=7B2026.05 | 36.29 | 53.06 | 52.62 | |
| GAP (LH+PCA)Latent Head (LH)=true, PCA alignment (PCA)=true2026.05 | 35.24 | 52 | 52.48 | |
| Monet-7BParameters=7B2026.05 | 32.67 | 50 | 47.99 | |
| GAP (LH+DA, no PCA)Latent Head (LH)=true, Difficulty-Aware supervision (DA)=true, PCA alignment (PCA)=false2026.05 | 31.26 | 50 | 50.05 | |
| GAP (LH)Latent Head (LH)=true2026.05 | 30 | 48.86 | 49.15 | |
| Dense Cap SFT2026.05 | 28.6 | 45.43 | 47.24 | |
| LVR2026.05 | 26.95 | 49.05 | 47.66 |