Models can often answer questions about a short clip, but a television drama unfolds across many episodes. Remembering who did what, and how one event led to another, gets harder when evidence is spread over many hours.
DramaVQA is a test set of questions about television dramas, from short sections to full series. Artificial intelligence drafts questions from video descriptions, then people check and correct them.
Models do reasonably well on shorter sections but fall to little better than guessing on the longest videos. The results show that simply giving models more video at once is not enough; they need ways to store and find important details over time.