Researchers have introduced Open-o3-Video-v2, an advanced agentic system designed for verifiable video reasoning. This model actively seeks and grounds supporting visual evidence, including timestamps and object-level bounding boxes, within its reasoning process. To facilitate this, new datasets like STGR-v2-CoT135K and STGR-v2-RL-13K have been created, alongside a novel tracking algorithm called Clue-Seek GRPO (CS-GRPO). Open-o3-Video-v2-4B has demonstrated significant performance improvements on benchmarks such as V-STaR, outperforming its base model and showing gains in general video understanding and temporal grounding. AI
IMPACT Enhances video reasoning capabilities by integrating active evidence acquisition and object-level grounding.
RANK_REASON The item describes a new research paper detailing a novel video reasoning model and associated datasets and algorithms. [lever_c_demoted from research: ic=1 ai=1.0]
Read on Hugging Face Daily Papers →
- Clue-Seek GRPO
- CS-GRPO
- Open-o3-Video
- Open-o3-Video-v2
- Open-o3-Video-v2-4B
- Qwen3-VL 4B
- STGR-v2-CoT135K
- STGR-v2-RL-13K
- V Star
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
