Research map / Video / Video-language understanding
Video-o3: Native Interleaved Clue Seeking for Long Video Multi-Hop Reasoning
Xiangyu Zeng, Zhiqiu Zhang, Yuhan Zhu, Xinhao Li, Zikang Wang, Changlian Ma, Qingyu Zhang, Zizheng Huang, Kun Ouyang, Tianxiang Jiang, Ziang Yan, Yi Wang and 3 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 |
|---|---|---|---|
| add task-decoupled attention masking in SFT (ablation) | Video-o3 (SFT) · LongVideoBench | accuracy: 55.0 -> 56.6 | Improved |
| add task-decoupled attention masking in SFT (ablation) | Video-o3 (SFT) · MMVU | accuracy: 60.6 -> 64.5 | Improved |
| add answer mask alone (vs full masking) (ablation) | Video-o3 (SFT) · LVBench | accuracy: 42.2 -> 44.2 | Improved |
| add verifiable trajectory-guided reward (ablation) | Video-o3 (RL) · Video-Holmes | accuracy: 42.9 -> 46.1 | Improved |
| add Hybrid Clue Score to reward (ablation) | Video-o3 (RL) · LVBench | accuracy: 45.2 -> 47.5 | Improved |
| add Turn Decay Factor to reward (ablation) | Video-o3 (RL) · Video-Holmes | accuracy: 45.5 -> 46.1 | Improved |
| raise max inference turns from 2 to 8 (ablation) | Video-o3 (SFT+RL) · MLVU | accuracy: 67.4 -> 72.1 | Improved |
| raise max inference turns from 2 to 8 (ablation) | Video-o3 (SFT+RL) · VideoMMMU | accuracy: 48.0 -> 51.7 | Improved |
| raise mask ratio from 0.1 to 0.3 (ablation) | Video-o3 (SFT) · Video-Holmes | accuracy: 41.3 -> 40.6 | Hurt |
| add Hybrid Clue Score (no grounding data) (ablation) | Video-o3 · Charades-STA | mIoU: 46.5 -> 53.8 | Improved |
| add Hybrid Clue Score (with grounding data) (ablation) | Video-o3 · Charades-STA | mIoU: 59.6 -> 60.7 | Improved |
| prompt Qwen2.5-VL zero-shot with tool use (ablation) | Qwen2.5-VL · LVBench | accuracy: 45.3 -> 42.8 | Hurt |
Later papers that compared with this method
2 later accepted papers report Video-o3 as a baseline in their results tables; 2 of them report a better number on most of the results they share with it.
- Towards Effective Long Video Understanding: Dynamic MAS Construction via Meta-Agent (ECCV 2026 · better on 4 of 4 shared results)
- VisionCoach: Reinforcing Grounded Video Reasoning via Visual-Perception Prompting (ECCV 2026 · better on 1 of 1 shared result)
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Video-Zoomer — VideoZoomer: Reinforcement-Learned Temporal Focusing for Long Video Reasoning (ICLR 2026) · better on 4 of 4 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Mini-o3: Scaling Up Reasoning Patterns and Interaction Turns for Visual Search (ICLR 2026 · similarity 0.90)
- Video-MTR: Reinforced Multi-Turn Reasoning for Long Video Understanding (ICML 2026 · similarity 0.90)
- LongVideo-R1: Smart Navigation for Low-cost Long Video Understanding (CVPR 2026 · similarity 0.89)
- FrameThinker: Learning to Think with Long Videos via Multi-Turn Frame Spotlighting (ICLR 2026 · similarity 0.89)
- TimeSearch-R: Adaptive Temporal Search for Long-Form Video Understanding via Self-Verification Reinforcement Learning (ICLR 2026 · similarity 0.89)
- When Thinking Drifts: Evidential Grounding for Robust Video Reasoning (NeurIPS 2025 · similarity 0.88)
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.
