Native Active Perception as Reasoning for Omni-Modal Understanding

arXiv·medium signal

Proposes replacing the passive 'watch-it-all' paradigm—where models process every frame uniformly regardless of the query—with native active perception that selectively attends to query-relevant content for long-video and omni-modal understanding. Frames perception itself as a reasoning step, which is practically relevant for cutting compute on long-context video pipelines. Authored by Zhenghao Xing, Ruiyang Xu, Yuxuan Wang et al. (cs.CV/cs.CL/cs.SD).