Reproduce “Open-o3-Video-v2: Seek, Ground, and Reason with Verifiable Spatio-Temporal Evidence” on OpenResearch

About this paper

Jiahao Meng, Zhongyuan Bao, Yue Tan, Haiquan Wen, Renrui Zhang, Anna Wang, Jason Li, Yanwei Fu and 2 more

Open-o3-Video-v2 combines iterative video seeking with timestamped object boxes to make retrieved evidence explicit in its reasoning. It requests intervals, inspects frames, and reuses localized evidence in later steps. On V-STaR, its 4B model scores 40.4% mAM and 56.3% mLGM, above its Qwen3-VL-4B backbone; this supports gains in the reported settings, not universal reliability.

Key insights

• Evidence retrieval and object-level grounding provide complementary benefits: multi-turn seeking improves results, and explicitly grounding acquired evidence improves them further.
• Rewarding evidence gained from each tool call—not just final answer correctness—helps the model search more selectively; matched-budget tests favor targeted retrieval over uniform or random frame selection.
• The model averages about 2.2 turns, including its final answer, and processes roughly 120 frames per question; remaining errors often involve temporal selection, object grounding, and fine-grained perception.

Summary by alphaXiv.