What it is
Uses ground-truth annotations as oracle rollouts in RL for video MLLMs, fixing mixed-group advantage inversion and cutting CoT-heavy training and inference cost. Instead of using annotations only to score samples, it inserts each annotation as an oracle rollout and decouples oracle guidance from on-policy advantage estimation, avoiding advantage inversion. Sign-balanced pruning keeps the oracle plus strongest positive/negative rollouts for cheaper updates.
Key results
- 2.2x SFT step time vs 4.9x for GRPO with CoT
- VSI-Bench 73.1 vs GPT-5 55.0 and Gemini-3-Pro 55.1
- Temporal mIoU 66.0 vs prior best 62.5; tracking AO 78.2 vs 73.0
Shows paired annotations can serve as direct RL optimization targets, making video MLLM post-training more sample-efficient than rollout-only group RL.