Research map / Video / Video-language understanding
SVI-Bench: A Dynamic Microworld for Strategic Video Intelligence
Yulu Pan, Han Yi, Seongsu Ha, Mohaiminul Islam, Benjamin Zhang, Lorenzo Torresani, Gedas Bertasius
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 |
|---|---|---|---|
| fine-tune LLaVA-Video-7B on SVI-Bench training data | LLaVA-Video-7B · T1 | avg score: +1.28 | Improved |
| fine-tune LLaVA-Video-7B on SVI-Bench training data | LLaVA-Video-7B · T2 | accuracy: +36.9% | Improved |
| fine-tune Qwen3-VL-8B on training data | Qwen3-VL-8B · T5 | accuracy: +7.9% (to 44.82) | Improved |
| replace video with ground-truth text (oracle) (ablation) | GPT-5.2 · T9 | accuracy: 4.6 -> 54.0 | Improved |
| replace video with ground-truth text (oracle) (ablation) | GPT-5 · T6 | factual accuracy: 71.99 -> 87.19 | Improved |
| replace video with ground-truth text (oracle) (ablation) | GPT-5.2 · T4 | score: 2.06 -> 2.46 | Improved |
| replace video with ground-truth text (oracle) (ablation) | GPT-5.2 · T5 | accuracy: 38.2 -> 41.9 | Improved |
| replace video with ground-truth text (oracle) (ablation) | GPT-5 · T6 | saliency: 7.1 -> 20.6 | Improved |
| fine-tune Wan 2.1 on SVI-Bench with structured conditions vs off-the-shelf MagicMotion | Wan 2.1 · T7 basketball | Video mIoU: 0.466 -> 0.513 | Improved |
| goal-conditioned adaptation of Wan framework vs MagicMotion | Wan-based reference · T8 | goal accuracy: 31.4 -> 50.2 | Improved |
| task-specific fine-tuning at perception level | fine-tuned models · higher pillars | performance | No clear effect |
| increase near-duplicate distractors in retrieval pool (ablation) | InternVideo2 · T3 | R@100: drops by nearly half | Hurt |
Closest papers by meaning
- VS-Bench: Evaluating VLMs for Strategic Abilities in Multi-Agent Environments (CVPR 2026 · similarity 0.87)
- SportR: A Benchmark for Multimodal Large Language Model Reasoning in Sports (ICLR 2026 · similarity 0.86)
- DynaMind: Reasoning over Abstract Video Dynamics for Embodied Decision-Making (ICML 2025 · similarity 0.86)
- SciVideoBench: Benchmarking Scientific Video Reasoning in Large Multimodal Models (ICML 2026 · similarity 0.85)
- VELOCITI: Benchmarking Video-Language Compositional Reasoning with Strict Entailment (CVPR 2025 · similarity 0.84)
- Event2Tracking: Reconstructing Multi-Agent Soccer Trajectories Using Long-Term Multimodal Context (AAAI 2025 · similarity 0.84)
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.
