Multimodal Reasoning on SLAKE, VQA-Rad, and Geo3K
68.87Overall PerformanceVISTA-SFT (Iteration 3)
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| VISTA-SFT (Iteration 3)Backbone=Qwen2.5-VL-7B-Instruct, |Train| Avg.=8.3K2026.05 | 68.87 | 87.89 | 77.29 | 41.43 | |
| VISTA-SFT (Iteration 2)Backbone=Qwen2.5-VL-7B-Instruct, |Train| Avg.=8.1K2026.05 | 67.12 | 84.79 | 75.3 | 41.26 | |
| R3V (Iteration 3)Backbone=Qwen2.5-VL-7B-Instruct, |Train| Avg.=18.7K2026.05 | 66.76 | 87.32 | 75.7 | 37.27 | |
| R3V (Iteration 2)Backbone=Qwen2.5-VL-7B-Instruct, |Train| Avg.=18.8K2026.05 | 66.74 | 86.48 | 75.3 | 38.44 | |
| VISTA-SFT (Iteration 1)Backbone=Qwen2.5-VL-7B-Instruct, |Train| Avg.=7.7K2026.05 | 66.44 | 83.66 | 74.9 | 40.77 | |
| ReSTEM (Iteration 2)Backbone=Qwen2.5-VL-7B-Instruct, |Train| Avg.=9.9K2026.05 | 66.35 | 84.79 | 74.49 | 39.77 | |
| ReSTEM (Iteration 3)Backbone=Qwen2.5-VL-7B-Instruct, |Train| Avg.=10.7K2026.05 | 66.25 | 84.51 | 75.3 | 38.94 | |
| RFTBackbone=Qwen2.5-VL-7B-Instruct, |Train| Avg.=25.5K2026.05 | 66 | 82.82 | 74.9 | 40.27 | |
| VISTA-SFT (Iteration 3)Backbone=Qwen2.5-VL-3B-Instruct, |Train| Avg.=8.3K2026.05 | 65.87 | 84.23 | 76.1 | 37.27 | |
| VISTA-SFT2026.05 | 65.3 | — | — | — | |
| R3V (Iteration 1)Backbone=Qwen2.5-VL-7B-Instruct, |Train| Avg.=18.5K2026.05 | 64.85 | 82.82 | 75.3 | 36.44 | |
| SFT-Oracle2026.05 | 64.63 | — | — | — | |
| SFT-OracleBackbone=Qwen2.5-VL-7B-Instruct, |Train| Avg.=1.3K2026.05 | 64.6 | 82.25 | 70.12 | 41.43 | |
| ReSTEM (Iteration 1)Backbone=Qwen2.5-VL-7B-Instruct, |Train| Avg.=8.5K2026.05 | 63.97 | 82.82 | 73.31 | 35.77 | |
| STaR (Iteration 3)Backbone=Qwen2.5-VL-7B-Instruct, |Train| Avg.=1.1K2026.05 | 63.68 | 84.23 | 69.72 | 37.1 | |
| VISTA-SFT (Iteration 2)Backbone=Qwen2.5-VL-3B-Instruct, |Train| Avg.=8.1K2026.05 | 63.39 | 83.38 | 72.51 | 34.28 | |
| STaR (Iteration 2)Backbone=Qwen2.5-VL-7B-Instruct, |Train| Avg.=1.0K2026.05 | 63.35 | 80 | 72.11 | 37.94 | |
| R3V (Iteration 2)Backbone=Qwen2.5-VL-3B-Instruct, |Train| Avg.=18.8K2026.05 | 63.18 | 82.82 | 74.1 | 32.61 | |
| R3V2026.05 | 63.15 | — | — | — | |
| STaR (Iteration 1)Backbone=Qwen2.5-VL-7B-Instruct, |Train| Avg.=0.9K2026.05 | 62.92 | 80 | 71.31 | 37.44 | |
| STaR+2026.05 | 62.92 | — | — | — | |
| SFT-OracleBackbone=Qwen2.5-VL-3B-Instruct, |Train| Avg.=1.3K2026.05 | 62.68 | 81.97 | 68.13 | 37.94 | |
| ReSTEM (Iteration 3)Backbone=Qwen2.5-VL-3B-Instruct, |Train| Avg.=10.7K2026.05 | 62.56 | 81.69 | 73.71 | 32.28 | |
| VISTA-SFT (Iteration 1)Backbone=Qwen2.5-VL-3B-Instruct, |Train| Avg.=7.7K2026.05 | 62.41 | 80.85 | 74.1 | 32.28 | |
| SFT-SeedBackbone=Qwen2.5-VL-7B-Instruct, |Train| Avg.=1.0K2026.05 | 62.2 | 79.15 | 70.52 | 36.94 | |
| R3V (Iteration 3)Backbone=Qwen2.5-VL-3B-Instruct, |Train| Avg.=18.7K2026.05 | 61.17 | 81.41 | 69.32 | 32.78 | |
| ReSTEM (Iteration 2)Backbone=Qwen2.5-VL-3B-Instruct, |Train| Avg.=9.9K2026.05 | 60.71 | 79.38 | 71.31 | 31.45 | |
| STaR2026.05 | 60.56 | — | — | — | |
| STaR (Iteration 3)Backbone=Qwen2.5-VL-3B-Instruct, |Train| Avg.=1.1K2026.05 | 60.13 | 80.28 | 69.32 | 30.78 | |
| RFTBackbone=Qwen2.5-VL-3B-Instruct, |Train| Avg.=25.5K2026.05 | 59.35 | 79.15 | 70.12 | 28.79 | |
| SFT-Seed2026.05 | 58.87 | — | — | — | |
| ReSTEM (Iteration 1)Backbone=Qwen2.5-VL-3B-Instruct, |Train| Avg.=8.5K2026.05 | 58.58 | 76.06 | 69.72 | 29.95 | |
| STaR (Iteration 2)Backbone=Qwen2.5-VL-3B-Instruct, |Train| Avg.=1.0K2026.05 | 58.06 | 76.9 | 65.34 | 31.95 | |
| R3V (Iteration 1)Backbone=Qwen2.5-VL-3B-Instruct, |Train| Avg.=18.5K2026.05 | 57.89 | 78.21 | 65.74 | 29.62 | |
| STaR (Iteration 1)Backbone=Qwen2.5-VL-3B-Instruct, |Train| Avg.=0.9K2026.05 | 57.35 | 73.8 | 66.13 | 32.11 | |
| SFT-SeedBackbone=Qwen2.5-VL-3B-Instruct, |Train| Avg.=1.0K2026.05 | 52.21 | 67.04 | 64.14 | 25.46 |