Multimodal Policy Internalization on ClevrPolicy-M N=6 (test)
85Accuracy (N=6)TriMPI w/ PoRo-DAPO
Evaluation results
| Method | Links | |
|---|---|---|
| TriMPI w/ PoRo-DAPOCPT=true, SFT=true, RL=true, RL Steps=90*|75*|50, PolicyRollout=true2025.10 | 85 | |
| TriMPI w/ PoRo-GRPOCPT=true, SFT=true, RL=true, RL Steps=104|104|50, PolicyRollout=true2025.10 | 84.7 | |
| TriMPI w/ DAPOCPT=true, SFT=true, RL=true, RL Steps=70*|70*|50, PolicyRollout=false2025.10 | 81.45 | |
| TriMPI w/ GRPOCPT=true, SFT=true, RL=true, RL Steps=104|104|50, PolicyRollout=false2025.10 | 80.8 | |
| CoT SFT + DAPOCPT=false, SFT=true, RL=true, RL Steps=104|85*|502025.10 | 74.4 | |
| CoT SFT + GRPOCPT=false, SFT=true, RL=true, RL Steps=104|104|502025.10 | 64.5 | |
| VM-CPT + CoT SFTCPT=true, SFT=true, RL=false2025.10 | 27.05 | |
| Direct SFTCPT=false, SFT=true, RL=false2025.10 | 14.55 | |
| CoT SFTCPT=false, SFT=true, RL=false2025.10 | 14.3 | |
| In-ContextMode=Zero-Shot with Policy In-Context2025.10 | 5.65 |