AI Paper Atlas

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

ECCV 2026 · Video › Video-language understanding · approach: Event and anomaly reasoning in video · 0 citations (OpenAlex)

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 changedWhereMeasuredEffect
fine-tune LLaVA-Video-7B on SVI-Bench training dataLLaVA-Video-7B · T1avg score: +1.28Improved
fine-tune LLaVA-Video-7B on SVI-Bench training dataLLaVA-Video-7B · T2accuracy: +36.9%Improved
fine-tune Qwen3-VL-8B on training dataQwen3-VL-8B · T5accuracy: +7.9% (to 44.82)Improved
replace video with ground-truth text (oracle) (ablation)GPT-5.2 · T9accuracy: 4.6 -> 54.0Improved
replace video with ground-truth text (oracle) (ablation)GPT-5 · T6factual accuracy: 71.99 -> 87.19Improved
replace video with ground-truth text (oracle) (ablation)GPT-5.2 · T4score: 2.06 -> 2.46Improved
replace video with ground-truth text (oracle) (ablation)GPT-5.2 · T5accuracy: 38.2 -> 41.9Improved
replace video with ground-truth text (oracle) (ablation)GPT-5 · T6saliency: 7.1 -> 20.6Improved
fine-tune Wan 2.1 on SVI-Bench with structured conditions vs off-the-shelf MagicMotionWan 2.1 · T7 basketballVideo mIoU: 0.466 -> 0.513Improved
goal-conditioned adaptation of Wan framework vs MagicMotionWan-based reference · T8goal accuracy: 31.4 -> 50.2Improved
task-specific fine-tuning at perception levelfine-tuned models · higher pillarsperformanceNo clear effect
increase near-duplicate distractors in retrieval pool (ablation)InternVideo2 · T3R@100: drops by nearly halfHurt

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy