Visual Policy Reasoning on ClevrPolicy-T (N=4)
99.15AccuracyTriMPI w/ PoRo-GRPO
Evaluation Results
| Method | Links | |
|---|---|---|
| TriMPI w/ PoRo-GRPOComplexity=N=42025.10 | 99.15 | |
| CoT SFT + GRPOComplexity=N=42025.10 | 98.25 | |
| TriMPI w/ PoRo-DAPOComplexity=N=42025.10 | 97.3 | |
| CoT SFT + DAPOComplexity=N=42025.10 | 97.15 | |
| CoT SFTComplexity=N=42025.10 | 49.1 | |
| In-ContextComplexity=N=42025.10 | 32.2 | |
| Direct SFTComplexity=N=42025.10 | 25.65 |