Visual Policy Reasoning on ClevrPolicy-T (N=6)
77.8AccuracyTriMPI w/ PoRo-DAPO
Evaluation Results
| Method | Links | |
|---|---|---|
| TriMPI w/ PoRo-DAPOModel Size=7B2025.10 | 77.8 | |
| TriMPI w/ PoRo-DAPOModel Size=3B2025.10 | 68.75 | |
| CoT SFT + DAPOModel Size=7B2025.10 | 67.6 | |
| TriMPI w/ PoRo-GRPOModel Size=7B2025.10 | 65.85 | |
| CoT SFT + DAPOModel Size=3B2025.10 | 48.1 | |
| CoT SFT + GRPOModel Size=7B2025.10 | 47.05 | |
| TriMPI w/ PoRo-GRPOModel Size=3B2025.10 | 46.65 | |
| CoT SFT + GRPOModel Size=3B2025.10 | 31.55 | |
| CoT SFTModel Size=3B2025.10 | 19 | |
| CoT SFTModel Size=7B2025.10 | 17.8 | |
| Direct SFTModel Size=7B2025.10 | 15.15 | |
| Direct SFTModel Size=3B2025.10 | 13.4 | |
| In-ContextModel Size=7B2025.10 | 13.15 | |
| In-ContextModel Size=3B2025.10 | 4.8 |