Research map / Video / Video-language understanding
VideoTemp-o3: Harmonizing Temporal Grounding and Video Understanding in Agentic Thinking-with-Videos
Wenqi Liu, Yunxiao Wang, Shijie Ma, Meng Liu, Qile Su, Tianke Zhang, Haonan fan, Changyi Liu, Kaiyu Jiang, Jiankang Chen, Kaiyu Tang, Bin Wen and 5 more
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply RL after SFT | VideoTemp-o3-7B · VideoMME Overall | accuracy: 60.6 -> 64.5 | Improved |
| add grounding data to SFT (ablation) | VideoTemp-o3 · ReXTime | mIoU: 13.0 -> 29.5 | Improved |
| add grounding data to SFT (ablation) | VideoTemp-o3 · VideoMME Overall | accuracy: 63.0 -> 64.5 | Improved |
| add unified masking (supervise only last two turns) (ablation) | VideoTemp-o3 SFT · VideoMMMU Overall | accuracy: 47.9 -> 53.2 | Improved |
| add unified masking (ablation) | VideoTemp-o3 SFT · ReXTime | mIoU: 18.8 -> 29.5 | Improved |
| add IoU reward in RL (ablation) | VideoTemp-o3 RL · VideoMMMU Overall | accuracy: 51.6 -> 53.2 | Improved |
| use penalty-aware IoU reward instead of plain IoU reward (ablation) | VideoTemp-o3 RL · VideoMMMU Overall | accuracy: 44.2 -> 53.2 | Improved |
| use penalty-aware IoU reward instead of plain IoU reward (ablation) | VideoTemp-o3 RL · ReXTime | mIoU: 23.8 -> 29.5 | Improved |
| learned on-demand cropping instead of always cropping once (ablation) | VideoTemp-o3 · VideoMME (w/o subtitle) Overall | accuracy: 63.0 -> 64.5 | Improved |
| learned tool use instead of answering directly without tool call (ablation) | VideoTemp-o3 · VideoMME (w/o subtitle) Overall | accuracy: 62.1 -> 64.5 | Improved |
| add prompt-only re-grounding to the base model | Qwen2.5-VL-7B · VideoMME (w/o subtitle) Overall | accuracy: 59.9 -> 58.4 | No clear effect |
| add inference-time self-correction to the base model | Qwen2.5-VL-7B · VideoMME (w/o subtitle) Overall | accuracy: 59.9 -> 60.5 | No clear effect |
Closest papers by meaning
- LongVT: Incentivizing "Thinking with Long Videos" via Native Tool Calling (CVPR 2026 · similarity 0.92)
- Thinking With Videos: Multimodal Tool-Augmented Reinforcement Learning for Long Video Reasoning (CVPR 2026 · similarity 0.90)
- Searching Videos as Trees: Self-Correcting Agents for Grounded Long Video QA (NeurIPS 2026 · similarity 0.89)
- Enhancing Video-LLM Reasoning via Agent-of-Thoughts Distillation (CVPR 2025 · similarity 0.88)
- Grounded Question-Answering in Long Egocentric Videos (CVPR 2024 · similarity 0.87)
- Temporal Chain of Thought: Long-Video Understanding by Thinking in Frames (NeurIPS 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
