Definition
Plain language
When an AI decides what to look at next instead of being fed everything at once.
As stated in the literature
A paradigm where a model treats acquiring observations (which frames, windows, or audio to sample) as actions chosen during reasoning rather than ingesting all input passively, letting compute scale with task difficulty instead of input length.
Also called: native active perception
Why it matters: It lets a system spend its effort only where the task is hard, so long videos or recordings don't force wasteful processing of irrelevant detail.
For example, when asked who scored the winning goal, an AI watching a two-hour match jumps straight to the final minutes instead of replaying every frame.