Research map / Video / Video-language understanding
Demystifing Video Reasoning
Ruisi Wang, Zhongang Cai, Fanyi Pu, Junxiang Xu, Wanqi Yin, Maijunxian Wang, Ran Ji, Chenyang Gu, Bo Li, Ziqi Huang, Hokin Deng, Dahua Lin and 2 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 Training-Free Ensemble (3 seeds, mid-layer latent averaging) | VBVR-Wan2.2 · VBVR-Bench overall | score: 0.685 -> 0.716 | Improved |
| add Training-Free Ensemble | VBVR-Wan2.2 · V-Reason Bench | score: 8.94 -> 12.12 | Improved |
| add Training-Free Ensemble | VBVR-LTX2.3 · V-Reason Bench | score: 3.61 -> 3.19 | Hurt |
| add Training-Free Ensemble | VBVR-Wan2.2 · RISE-Video | score: 61.60 -> 65.85 | Improved |
| inject noise at a diffusion step (all frames) | VBVR-Wan2.2 · VBVR-Bench | final score: 0.685 to below 0.3 | Hurt |
| inject noise in a single frame across all steps | VBVR-Wan2.2 · VBVR-Bench | final score | No clear effect |
| aggregate layers 20-29 instead of 0-9 (ablation) | TFE · VBVR-Bench overall | score: 0.688 -> 0.716 | Improved |
| aggregate layers 20-29 instead of all layers 0-39 (ablation) | TFE · VBVR-Bench overall | score: 0.690 -> 0.716 | Improved |
| increase ensemble size from 3 to 4 seeds (ablation) | TFE · VBVR-Bench overall | score: 0.716 -> 0.729 | Improved |
| increase ensemble size from 3 to 6 seeds (ablation) | TFE · VBVR-Bench overall | score: 0.716 -> 0.675 | Hurt |
| use multi-frame generation (~100 frames) instead of single-frame ChronoEdit variant (ablation) | VBVR-Wan2.2 · VBVR-Bench overall | score: 0.581 -> 0.685 | Improved |
| reduce frames from ~100 to 5 (ablation) | VBVR-Wan2.2 · VBVR-Bench overall | score: 0.685 -> 0.619 | Hurt |
Closest papers by meaning
- When Thinking Drifts: Evidential Grounding for Robust Video Reasoning (NeurIPS 2025 · similarity 0.92)
- OpenCoF: Learning to Reason Through Video Generation (NeurIPS 2026 · similarity 0.91)
- Video-of-Thought: Step-by-Step Video Reasoning from Perception to Cognition (ICML 2024 · similarity 0.91)
- Look, Remember and Reason: Grounded Reasoning in Videos with Language Models (ICLR 2024 · similarity 0.90)
- CollabVR: Collaborative Video Reasoning with Vision-Language and Video Generation Models (NeurIPS 2026 · similarity 0.90)
- VideoRFT: Incentivizing Video Reasoning Capability in MLLMs via Reinforced Fine-Tuning (NeurIPS 2025 · 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.
