Answer first, check later: AI cuts video-search cost by 93%

Ask an AI a question about a three-hour video — "When did the person take their medication?" — and you hit a brutal trade-off. Either you feed every frame to a powerful but expensive vision model (slow and costly), or you skim short text descriptions of the footage (cheap, but you miss the quick physical moments that actually answer the question).

TimeProVe splits the work into two stages: guess cheap, then check expensive.

First, lightweight tools scan the video for individual actions and pin each one to a time stamp — washing hands at 14:02, opening a cabinet at 14:05. A small, cheap language model then reads those time-stamped actions and proposes likely answers, each paired with the short stretch of video that supports it. Think of it as a junior assistant flagging "the answer is probably here, around minute 14."

Only then does the costly vision model get called in — not to watch the whole video, but to verify those few flagged moments. That's the savings: 75% fewer calls to the expensive model and 93% lower running cost, while still beating the best existing method by 7.3 points on the authors' new test set of everyday home activities.

The gains lean on the action-detector catching the right moments up front; if it misses a subtle action, the verifier never gets a chance to confirm it. And the headline numbers come from a benchmark the same team designed.

Why you should care: Searching long footage for one fleeting moment usually means paying to process hours you don't need. This propose-then-verify split makes that search roughly fourteen times cheaper.