Pass@4 Reasoning on Dyna Math
89.4Pass@4Base
Evaluation results
| Method | Links | |
|---|---|---|
| BaseModel Size=32B, Backbone=Qwen3-VL-Thinking2026.05 | 89.4 | |
| AXPOModel Size=4B, Backbone=Qwen3-VL-Thinking2026.05 | 87 | |
| SFT + GRPOModel Size=8B, Backbone=Qwen3-VL-Thinking2026.05 | 87 | |
| AXPOModel Size=8B, Backbone=Qwen3-VL-Thinking2026.05 | 87 | |
| SFTModel Size=4B, Backbone=Qwen3-VL-Thinking2026.05 | 86.4 | |
| SFTModel Size=8B, Backbone=Qwen3-VL-Thinking2026.05 | 86 | |
| SFT + GRPOModel Size=4B, Backbone=Qwen3-VL-Thinking2026.05 | 85.8 | |
| BaseModel Size=8B, Backbone=Qwen3-VL-Thinking2026.05 | 85.2 | |
| GRPOModel Size=8B, Backbone=Qwen3-VL-Thinking2026.05 | 84.6 | |
| GRPOModel Size=4B, Backbone=Qwen3-VL-Thinking2026.05 | 84.2 | |
| AXPOModel Size=2B, Backbone=Qwen3-VL-Thinking2026.05 | 83.6 | |
| BaseModel Size=4B, Backbone=Qwen3-VL-Thinking2026.05 | 82.6 | |
| SFTModel Size=2B, Backbone=Qwen3-VL-Thinking2026.05 | 80.2 | |
| SFT + GRPOModel Size=2B, Backbone=Qwen3-VL-Thinking2026.05 | 80.2 | |
| GRPOModel Size=2B, Backbone=Qwen3-VL-Thinking2026.05 | 78.2 | |
| BaseModel Size=2B, Backbone=Qwen3-VL-Thinking2026.05 | 76.6 |