Visual Mathematical Reasoning on MathVista (Accuracy, Latency, Speedup)
75.4AccuracyDRL-trained M_d_dft (7B)
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| DRL-trained M_d_dft (7B)Draft (1 - pt)%=20%2026.06 | 75.4 | 55.5 | 1.17 | |
| DRL-trained M_d_dft (7B)Draft (1 - pt)%=30%2026.06 | 75.3 | 53.5 | 1.21 | |
| JRL-trained M_d_joint (7B)Draft (1 - pt)%=30%2026.06 | 74.4 | 47.6 | 1.36 | |
| DRL-trained M_d_dft (7B)Draft (1 - pt)%=50%2026.06 | 74.3 | 43.6 | 1.49 | |
| JRL-trained M_d_joint (7B)Draft (1 - pt)%=60%2026.06 | 74.2 | 26.9 | 2.41 | |
| OursDraft (1 - pt)%=60%2026.06 | 74.2 | 26.9 | 2.41 | |
| Mt(Skywork-R1V-38B)2026.06 | 74.1 | 64.8 | — | |
| JRL-trained M_d_joint (7B)Draft (1 - pt)%=40%2026.06 | 74.1 | 40.1 | 1.62 | |
| JRL-trained M_d_joint (7B)Draft (1 - pt)%=50%2026.06 | 74 | 32.6 | 1.99 | |
| Token-probability-based (Paradigm 1)Draft (1 - pt)%=60%2026.06 | 72.2 | 29.5 | 2.19 | |
| Special-token-based (Paradigm 2)Draft (1 - pt)%=60%2026.06 | 72.1 | 26.9 | 2.41 | |
| RandomDraft (1 - pt)%=60%2026.06 | 71.6 | 27.5 | 2.36 | |
| JRL-trained M_d_joint (7B)Draft (1 - pt)%=100%2026.06 | 70 | 4.9 | — | |
| DRL-trained M_d_dft (7B)Draft (1 - pt)%=100%2026.06 | 69.9 | 10.2 | — |