VidSum-Reason

We present VidSum-Reason, a new benchmark for fine-grained, reasoning-driven video summarization from natural-language queries.Unlike prior datasets whose queries are short, literal keywords, VidSum-Reason poses 20 diverse video\u2013query pairs spanning four increasingly challenging categories\u2014standard, attribute-aware, temporal reasoning, and reasoning with external knowledge.The corpus comprises 9 videos (2\u20136 min) drawn from sports, DIY tutorials, movie trailers, CGI shorts, and YouTube-8M clips, offering rich visual and semantic variety.Each video is segmented into 2 s clips and scored on a 1\u20135 scale for both query relevance and overall narrative importance, with labels vetted by multiple annotators to ensure consistency.Evaluation follows the key-shot F1 protocol with a 36 % budget, enabling direct comparison to SumMe and TVSum while exposing the limitations of keyword-matching systems.VidSum-Reason thus provides the first testbed that jointly probes intent interpretation, temporal reasoning, and real-world knowledge, catalyzing research on genuinely query-focused, knowledge-aware video summarization.