AI Video Tool Now Answers Faster and Cheaper Than Ever

Streaming video is hard: questions arrive at any moment over an endless feed, and for StreamScout, deciding how deeply to search memory matters just as much as what to store. StreamScout keeps only a lightweight text timeline in context, then progressively adds up to three more detailed visual views—recent frames, a uniform look-back, and query-salient retrieval—until it can answer with confidence; if not, it escalates to the next view. Two trained variants, StreamScout-S and StreamScout-R, sharpen this stop-or-escalate policy using distillation and reinforcement learning. Across three backbones and three streaming benchmarks, StreamScout outperforms prior streaming methods while substantially cutting inference cost and token consumption. On OVO-Bench specifically, StreamScout-S improves Qwen3-VL-8B by 14.65 points, uses 59% fewer tokens than uniform sampling, and answers in 1.04 seconds on average.