Goal. For each question, retrieve and temporally ground the salient segment(s) across a pool of videos and return a concise, step‑wise answer linked to those segments. Models must reason over multiple videos and use audio + visual cues.

Why this matters. Most video QA benchmarks assume a single clip. Real‑world search requires cross‑video retrieval, timestamp grounding, and multi‑step reasoning.



Baseline. A simple retrieval‑then‑extract pipeline using transcripts + timestamps is provided.

Tracks.

Main: Audio‑Visual (both audio and visual features allowed)

Audio‑Only: (optional secondary leaderboard)