Open-o3-Video-v2: Seek, Ground, and Reason with Verifiable Spatio-Temporal Evidence
Open-o3-Video-v2 adds active evidence seeking to video reasoning, letting the model fetch missing frames before answering.
The paper extends Open-o3-Video beyond a fixed sparse frame sampler. Its reasoning traces can include timestamps and object-level boxes, with the model deciding when it needs more visual evidence. The authors introduce STGR-v2-CoT-135K, STGR-v2-RL-13K, and CS-GRPO to train and reward that behavior. On V-STaR, the 4B model beats its Qwen3-VL-4B backbone by 8.5 points mAM and 12.6 points mLGM. Source: HF Daily Papers' note.
The paper extends Open-o3-Video beyond a fixed sparse frame sampler. Its reasoning traces can include timestamps and object-level boxes, with the model deciding when it needs more visual evidence. The authors introduce STGR-v2-CoT-135K, STGR-v2-RL-13K, and CS-GRPO to train and reward that behavior. On V-STaR, the 4B model beats its Qwen3-VL-4B backbone by 8.5 points mAM and 12.6 points mLGM. Source: HF Daily Papers' note.
score 5