VBVR
A Very Big Video Reasoning Suite — benchmarking reasoning in video models at scale. ICML 2026.
Video Reasoning Benchmark · Northeastern University · Sept 2025 – Jan 2026 · ICML 2026
Paper · Code · Website · Dataset
- Proposed and implemented VBVR-EvalKit, a benchmark for evaluating reasoning capabilities in video models, covering 40 models.
- Co-led a team of 22 researchers: designed the data generation and benchmark framework, defined the development roadmap, and contributed core code and evaluation metrics. Launched an open-source ecosystem including 100+ synthetic data generators.
- Proposed and built a video reasoning dataset covering 200 curated reasoning tasks and 1M+ video clips — 1000× larger than prior datasets.