Video Question Answering on Video-MME short and medium videos
50.1Temporal AccuracyOracle-RLAIF
Evaluation Results
| Method | Links | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Oracle-RLAIFOptimization=Rank-based RLAIF2025.10 | 50.1 | 38.1 | 49 | 43.4 | 42.3 | 29.5 | 37.8 | 48 | 43.2 | |
| Oracle-DPOOptimization=DPO2025.10 | 49.5 | 34.8 | 46.3 | 41.4 | 40.2 | 27.1 | 38.1 | 48.5 | 41.4 | |
| VLM SFTOptimization=Supervised Fine-Tuning2025.10 | 48.2 | 32.9 | 39.5 | 43.3 | 39.5 | 19.6 | 34.5 | 50.1 | 37.5 | |
| VLM-RLAIFOptimization=RLAIF (scalar reward + PPO)2025.10 | 38.8 | 29.6 | 36.8 | 45.1 | 39.4 | 19.4 | 34.6 | 49.4 | 36.9 |