Research map / Video / Video-language understanding
VideoKR: Towards Knowledge- and Reasoning-Intensive Video Understanding
Lin Fu, Zheyuan Yang, Yang Wang, Tingyu Song, Arman Cohan, Yilun Zhao
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 the knowledge-intensive reasoning skill (KnowVidR) to VidR + KnowVid SFT data (ablation) | Qwen2.5-VL-7B SFT · Knowledge-intensive average | accuracy: 41.3 -> 42.4 | Improved |
| add KnowVid to VidR only SFT data (ablation) | Qwen2.5-VL-7B SFT · Knowledge-intensive average | accuracy: 41.4 -> 41.3 | No clear effect |
| add KnowVid to VidR only SFT data (ablation) | Qwen2.5-VL-7B SFT · VideoKR-Eval | accuracy: 35.3 -> 35.9 | Improved |
| add KnowVidR to VidR + KnowVid (ablation) | Qwen2.5-VL-7B SFT · VideoKR-Eval | accuracy: 35.9 -> 36.8 | Improved |
| add chain-of-thought rationales to SFT data instead of direct output (ablation) | Qwen2.5-VL-7B SFT · Knowledge-intensive average | accuracy: 39.4 -> 42.4 | Improved |
| add chain-of-thought rationales to SFT data instead of direct output (ablation) | Qwen2.5-VL-7B SFT · General video reasoning average | accuracy: 61.4 -> 58.3 | Hurt |
| train on VideoKR-SFT-201K instead of Video-R1-CoT-165k | Qwen2.5-VL-7B SFT · Knowledge-intensive average | accuracy: 36.2 -> 42.4 | Improved |
| train on VideoKR-SFT-201K instead of VideoRFT-CoT-102K | Qwen2.5-VL-7B SFT · Knowledge-intensive average | accuracy: 38.4 -> 42.4 | Improved |
| GRPO on VideoKR-RL-114K instead of VideoAuto-R1-83K | Qwen2.5-VL-7B zero-RL · Knowledge-intensive average | accuracy: 42.7 -> 43.0 | Improved |
| add RL after SFT on VideoKR (ablation) | Qwen3-VL-8B · Knowledge-intensive average | accuracy: 49.2 -> 51.5 | Improved |
| zero RL only versus SFT only on VideoKR (ablation) | Qwen3-VL-8B · Knowledge-intensive average | accuracy: 49.2 -> 50.6 | Improved |
| post-train Qwen2.5-VL-7B with VideoKR SFT + RL | Qwen2.5-VL-7B · Knowledge-intensive average | accuracy: 41.9 -> 46.6 (+4.7) | Improved |
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.
- VideoAuto-R1 — VideoAuto-R1: Video Auto Reasoning via Thinking Once, Answering Twice (CVPR 2026) · better on 2 of 3 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
- VideoEspresso: A Large-Scale Chain-of-Thought Dataset for Fine-Grained Video Reasoning via Core Frame Selection (CVPR 2025 · similarity 0.91)
- Video-R1: Reinforcing Video Reasoning in MLLMs (NeurIPS 2025 · similarity 0.90)
- Look, Remember and Reason: Grounded Reasoning in Videos with Language Models (ICLR 2024 · similarity 0.90)
- OpenCoF: Learning to Reason Through Video Generation (NeurIPS 2026 · similarity 0.89)
- VideoPrism: A Foundational Visual Encoder for Video Understanding (ICML 2024 · similarity 0.89)
- VideoReasonBench: Can MLLMs Perform Vision-Centric Complex Video Reasoning? (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.
