arXiv 2606.20561 introduces TimeProVe, a "propose-then-verify" architecture for long-video question answering. The standout: instead of having the VLM "watch the whole video" (full audit), TimeProVe first has a "proposer" model identify the relevant time segment, then a "verifier" model only attends to that segment. The result: 5-10× speedup with no quality loss.
The "propose-then-verify" pattern: the proposer is a lightweight model (1B parameters) that scans the video at low resolution and outputs a "time proposal" (e.g., "the relevant action happens between 12:34 and 12:51"). The verifier is a high-resolution VLM (7B parameters) that only watches the proposed time segment and answers the question.
The efficiency win: the proposer's low-resolution scan is O(n) in video length, and the verifier's high-resolution attention is only O(k) where k is the segment length (typically 1-5% of the total). The total compute is dominated by the proposer's scan, which is much cheaper.
The benchmark: on the long-video QA benchmark (1-hour videos), TimeProVe-7B-Verifier + 1B-Proposer matches the accuracy of a 7B VLM watching the full video, at 5-10× lower compute. The speedup is more dramatic for longer videos (10× for 1-hour, 15× for 3-hour).
The bigger takeaway: "propose-then-verify" is the right architecture for long-video understanding. The "watch everything" approach is wasteful, and the "find then focus" approach scales much better. For the industry, this means long-video products (video search, surveillance analysis, video summarization) can be deployed at much lower cost, opening up new use cases.