Visual Policy Reasoning on ClevrPolicy-M N=4
99.4AccuracyTriMPI w/ PoRo-DAPO
Evaluation Results
| Method | Links | |
|---|---|---|
| TriMPI w/ PoRo-DAPOComplexity=N=42025.10 | 99.4 | |
| TriMPI w/ PoRo-GRPOComplexity=N=42025.10 | 99.05 | |
| CoT SFT + DAPOComplexity=N=42025.10 | 88.75 | |
| CoT SFT + GRPOComplexity=N=42025.10 | 87.1 | |
| CoT SFTComplexity=N=42025.10 | 55.35 | |
| Direct SFTComplexity=N=42025.10 | 37.85 | |
| In-ContextComplexity=N=42025.10 | 13.9 |