Multimodal Policy Internalization on GTAPolicy (test)
85.85Tool AccuracyTriMPI w/ PoRo-GRPO
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| TriMPI w/ PoRo-GRPOCPT=true, SFT=true, RL=true, RL Steps=104|104|50, PolicyRollout=true2025.10 | 85.85 | 76.28 | 81.06 | |
| TriMPI w/ GRPOCPT=true, SFT=true, RL=true, RL Steps=104|104|50, PolicyRollout=false2025.10 | 83.96 | 74.7 | 79.33 | |
| TriMPI w/ PoRo-DAPOCPT=true, SFT=true, RL=true, RL Steps=90*|75*|50, PolicyRollout=true2025.10 | 80.19 | 71.82 | 76.01 | |
| TriMPI w/ DAPOCPT=true, SFT=true, RL=true, RL Steps=70*|70*|50, PolicyRollout=false2025.10 | 79.25 | 70.85 | 75.05 | |
| CoT SFT + GRPOCPT=false, SFT=true, RL=true, RL Steps=104|104|502025.10 | 76.42 | 68.02 | 72.22 | |
| CoT SFT + DAPOCPT=false, SFT=true, RL=true, RL Steps=104|85*|502025.10 | 76.42 | 68.44 | 72.43 | |
| VM-CPT + CoT SFTCPT=true, SFT=true, RL=false2025.10 | 69.81 | 61.13 | 65.47 | |
| CoT SFTCPT=false, SFT=true, RL=false2025.10 | 57.55 | 51.45 | 54.5 | |
| Direct SFTCPT=false, SFT=true, RL=false2025.10 | 44.34 | 37.16 | 40.75 | |
| In-ContextMode=Zero-Shot with Policy In-Context2025.10 | 23.58 | 19.44 | 21.51 |