Perception on Visual Probe
45.8Pass@1SFT + AXPO
Evaluation results
| Method | Links | |
|---|---|---|
| SFT + AXPOBackbone=Qwen3-VL-Thinking, Model Size=8B2026.05 | 45.8 | |
| SFT + AXPOBackbone=Qwen3-VL-Thinking, Model Size=4B2026.05 | 43.6 | |
| BaseBackbone=Qwen3-VL-Thinking, Model Size=32B2026.05 | 40.3 | |
| GRPOBackbone=Qwen3-VL-Thinking, Model Size=8B2026.05 | 40.1 | |
| SFT + GRPOBackbone=Qwen3-VL-Thinking, Model Size=8B2026.05 | 40.1 | |
| SFTBackbone=Qwen3-VL-Thinking, Model Size=8B2026.05 | 38.4 | |
| SFT + GRPOBackbone=Qwen3-VL-Thinking, Model Size=4B2026.05 | 38 | |
| SFT + GRPOBackbone=Qwen3-VL-Thinking, Model Size=2B2026.05 | 36.1 | |
| SFT + AXPOBackbone=Qwen3-VL-Thinking, Model Size=2B2026.05 | 35.1 | |
| SFTBackbone=Qwen3-VL-Thinking, Model Size=4B2026.05 | 34.7 | |
| BaseBackbone=Qwen3-VL-Thinking, Model Size=8B2026.05 | 31.8 | |
| GRPOBackbone=Qwen3-VL-Thinking, Model Size=4B2026.05 | 29.7 | |
| BaseBackbone=Qwen3-VL-Thinking, Model Size=4B2026.05 | 24.8 | |
| SFTBackbone=Qwen3-VL-Thinking, Model Size=2B2026.05 | 24.5 | |
| GRPOBackbone=Qwen3-VL-Thinking, Model Size=2B2026.05 | 23.8 | |
| BaseBackbone=Qwen3-VL-Thinking, Model Size=2B2026.05 | 5 |