Multi-modal reasoning on LLMs-Eval mini (test)
52.77GeoQAtest AccuracySAPO + AT-RL
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| SAPO + AT-RLBackbone=Qwen2.5-VL-7B-Instruct, Temperature=0.1, Credit Assignment=Anchor-RL2026.02 | 52.77 | 48.34 | 30.33 | 68.6 | 66.23 | 53.25 | |
| GSPO + AT-RLBackbone=Qwen2.5-VL-7B-Instruct, Temperature=0.1, Credit Assignment=Anchor-RL2026.02 | 50.74 | 47.46 | 29.32 | 68.1 | 64.46 | 52.02 | |
| DAPO + AT-RLBackbone=Qwen2.5-VL-7B-Instruct, Temperature=0.1, Credit Assignment=Anchor-RL2026.02 | 50.13 | 47.63 | 28.73 | 68.1 | 63.56 | 51.63 | |
| SAPOBackbone=Qwen2.5-VL-7B-Instruct, Temperature=0.1, Credit Assignment=Uniform2026.02 | 49.46 | 44.78 | 27.63 | 66 | 61.82 | 49.94 | |
| GRPO + AT-RLBackbone=Qwen2.5-VL-7B-Instruct, Temperature=0.1, Credit Assignment=Anchor-RL2026.02 | 49.24 | 45.98 | 27.97 | 66.7 | 62.73 | 50.52 | |
| VPPOBackbone=Qwen2.5-VL-7B-Instruct, Temperature=0.1, Credit Assignment=Fine-Grained2026.02 | 48.89 | 45.12 | 28.03 | 65.8 | 60.97 | 49.76 | |
| GSPOBackbone=Qwen2.5-VL-7B-Instruct, Temperature=0.1, Credit Assignment=Uniform2026.02 | 47.64 | 44.27 | 25.2 | 66.4 | 59.15 | 48.53 | |
| DAPOBackbone=Qwen2.5-VL-7B-Instruct, Temperature=0.1, Credit Assignment=Uniform2026.02 | 47.54 | 42.61 | 25.36 | 64.3 | 60.41 | 48.04 | |
| GRPOBackbone=Qwen2.5-VL-7B-Instruct, Temperature=0.1, Credit Assignment=Uniform2026.02 | 46.86 | 41.81 | 24.37 | 63.8 | 59.71 | 47.31 | |
| FT-RLBackbone=Qwen2.5-VL-7B-Instruct, Temperature=0.1, Credit Assignment=Fine-Grained2026.02 | 44.19 | 39.14 | 22.37 | 61.2 | 57.35 | 44.85 | |
| Zero-shotBackbone=Qwen2.5-VL-7B-Instruct, Temperature=0.12026.02 | 44.05 | 39.62 | 22.7 | 61.1 | 57.59 | 45.01 | |
| StepGRPOBackbone=Qwen2.5-VL-7B-Instruct, Temperature=0.1, Credit Assignment=Fine-Grained2026.02 | 41.12 | 35.34 | 19.8 | 59.7 | 55.23 | 42.24 |