Research map / Video / Video-language understanding
Are Video Reasoning Models Ready to Go Outside?
Yangfan He, Changgyu Boo, Jaehong Yoon
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 ROVA training to Qwen2.5-VL-7B | Qwen2.5-VL-7B · PVRBench (perturbed, avg) | accuracy: .33 -> .47 | Improved |
| add ROVA training to Embodied-R | Embodied-R 7B · PVRBench (perturbed, avg) | accuracy: 0.42 -> 0.50 | Improved |
| add ROVA training to Qwen2.5-VL-72B | Qwen2.5-VL-72B · PVRBench (perturbed, avg) | accuracy: .45 -> .56 | Improved |
| add ROVA training (effect on clean data) | Qwen2.5-VL-7B · PVRBench (clean, Orig.) | accuracy: .51 -> .53 | Improved |
| add curriculum (SRE+DRE+ME) to dual-branch alignment (ablation) | ROVA · PVRBench (avg acc, training efficiency) | GPU-hours and accuracy: 142.8 -> 134.4 GPU-h; .48 -> .53 (text: 0.37 to 0.47) | Improved |
| add dual-branch structured corruption alignment over standard GRPO (ablation) | Qwen2.5-VL-7B · PVRBench (avg acc, training efficiency) | accuracy: .45 -> .48 | Improved |
| difficulty-aware discarding instead of random discarding (ablation) | ROVA · PVRBench | accuracy: up to +3.4% | Improved |
| replace GPT-4o reward judge with open-source Qwen3-13B (ablation) | ROVA · PVRBench | accuracy: 0.470 -> 0.467 | No clear effect |
| use step-level or conditional alignment reward (ablation) | ROVA · PVRBench | accuracy | Hurt |
| structured perturbation-aware masks instead of random masking (ablation) | ROVA · PVRBench | accuracy: 6-9% absolute | Improved |
| increase perturbation intensity from 0.7 to 0.9 (ablation) | ROVA · PVRBench val | avg accuracy (%): 39.1 -> 36.8 | Hurt |
| lower perturbation intensity to 0.5 (ablation) | ROVA · PVRBench val | avg accuracy (%): 39.1 -> 40.2 | Mixed |
Closest papers by meaning
- MVVBench: Benchmarking 4D Reasoning in Vision-Language Models (NeurIPS 2026 · similarity 0.88)
- Black Swan: Abductive and Defeasible Video Reasoning in Unpredictable Events (CVPR 2025 · similarity 0.87)
- Video-R1: Reinforcing Video Reasoning in MLLMs (NeurIPS 2025 · similarity 0.87)
- SciVideoBench: Benchmarking Scientific Video Reasoning in Large Multimodal Models (ICML 2026 · similarity 0.86)
- GTR-Bench: Evaluating Geo-Temporal Reasoning in Vision-Language Models (ICLR 2026 · similarity 0.86)
- CollabVR: Collaborative Video Reasoning with Vision-Language and Video Generation Models (NeurIPS 2026 · similarity 0.86)
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.
