Paper summary: A Very Big Video Reasoning Suite

Problem

Video-model progress has focused on visual quality. Video reasoning (continuity, interaction, causality) is under-explored, partly because there’s no large-scale training data for it.

Method

• VBVR Dataset: 200 curated reasoning tasks following a taxonomy, with over one million clips, about three orders of magnitude larger than existing datasets.
• VBVR-Bench: a verifiable evaluation framework with rule-based, human-aligned scorers, not model judges, for reproducible, interpretable diagnosis.
• A large-scale scaling study of video reasoning.

Key results

From the abstract: one of the first large-scale scaling studies of video reasoning, with early signs of emergent generalisation to unseen reasoning tasks. Data, toolkit and models are released.

Limitations

There’s no dedicated limitations section in the extracted text. The authors describe the generalisation as “early signs”, not an established result.