Research map / Video / Video-language understanding
Video-KTR: Reinforcing Video Reasoning via Key Token Attribution
Ziyue Wang, Sheng Jin, ZHONGRONG ZUO, Jiawei Wu, Han Qiu, Qi She, Hao Zhang, Xudong Jiang
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 visual-aware token selection (ablation) | Vanilla GRPO · Video-Holmes/VideoMMMU/MMVU | average accuracy: 51.1 -> 52.5 | Improved |
| add temporal-aware token selection (ablation) | Vanilla GRPO · Video-Holmes/VideoMMMU/MMVU | average accuracy: 51.1 -> 52.6 | Improved |
| add entropy-aware token selection alone (ablation) | Vanilla GRPO · Video-Holmes/VideoMMMU/MMVU | average accuracy: 51.1 -> 51.0 | No clear effect |
| add temporal-aware tokens alone (ablation) | Vanilla GRPO · VideoMMMU | accuracy: 49.8 -> 50.1 | Mixed |
| add entropy tokens to visual and temporal (ablation) | V+T · Video-Holmes/VideoMMMU/MMVU | average accuracy: 52.9 -> 53.4 | Improved |
| add temporal tokens to visual and entropy (ablation) | V+E · Video-Holmes/VideoMMMU/MMVU | average accuracy: 52.8 -> 53.4 | Improved |
| use hard top-20% token selection instead of soft weighting (ablation) | Video-KTR · Video benchmarks | accuracy | Improved |
| use log-probability difference instead of distribution distances (ablation) | Video-KTR · Video benchmarks | accuracy | Improved |
| set token update ratio to 20% instead of 30-50% (ablation) | Video-KTR · Video benchmarks | accuracy | Improved |
| vary perturbation strategy (mask half, reversal, segmental shuffle) (ablation) | Video-KTR · Video benchmarks | average accuracy: 52.5; 52.6 | No clear effect |
| increase input frames from 16 to 64 | Video-KTR · Video-Holmes | accuracy: 40.7 -> 42.7 | Improved |
| replace vanilla GRPO with Video-KTR | Vanilla GRPO · MMMU(val) | accuracy: 44.33 -> 47.33 | Improved |
Closest papers by meaning
- Video-R1: Reinforcing Video Reasoning in MLLMs (NeurIPS 2025 · similarity 0.91)
- Incentivizing Versatile Video Reasoning in MLLMs via Data-Efficient Reinforcement Learning (CVPR 2026 · similarity 0.91)
- When Thinking Drifts: Evidential Grounding for Robust Video Reasoning (NeurIPS 2025 · similarity 0.90)
- Find, Fix, Reason: Context Repair for Video Reasoning (ICML 2026 · similarity 0.90)
- VideoRFT: Incentivizing Video Reasoning Capability in MLLMs via Reinforced Fine-Tuning (NeurIPS 2025 · similarity 0.90)
- Spotlight on Token Perception for Multimodal Reinforcement Learning (ICLR 2026 · similarity 0.89)
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.
