Active Video Perception: Iterative Evidence Seeking for Agentic Long Video Understanding

Researchers introduce Active Video Perception (AVP), an AI framework that enables agents to actively seek relevant evidence in long videos rather than passively processing entire content. The system uses an iterative plan-observe-reflect process to achieve superior accuracy on five benchmarks while reducing inference time by 82% and token usage by 88% compared to existing agentic methods.

Active Video Perception addresses a fundamental efficiency problem in video AI: current systems waste computational resources analyzing irrelevant content. Traditional approaches rely on query-agnostic video captioners that process entire videos regardless of what information matters for answering specific questions. AVP inverts this paradigm by implementing active perception—a cognitive science principle where observers intelligently direct attention toward task-relevant information rather than passively consuming everything available.

The framework's innovation lies in its iterative architecture. A planner proposes targeted interactions with video content, an observer extracts time-stamped evidence from specific temporal and spatial regions, and a reflector evaluates whether sufficient evidence exists to answer the query. This mirrors human reasoning: we don't watch entire videos to answer questions; we seek specific moments and details. The 5.7% accuracy improvement over previous agentic methods while using 82% less inference time and 88% fewer tokens represents a significant leap in efficiency.

For the AI industry, this research demonstrates practical progress toward cost-effective multimodal reasoning. Long video understanding powers real applications in surveillance, content moderation, video search, and documentary analysis—domains where computational efficiency directly impacts deployment viability. The dramatic reduction in token consumption is particularly valuable given the rising costs of large language model inference.

Looking forward, active perception principles could extend beyond video to other multimodal domains like document understanding and image analysis. The framework's success suggests that agency and selective attention, not just scale, drive capability gains in foundation models.

• →AVP achieves 5.7% higher accuracy than previous best agentic methods while using only 18.4% of inference time
• →The system iteratively plans observations, extracts evidence, and reflects on sufficiency rather than processing entire videos passively
• →Input token usage drops 87.6%, significantly reducing computational and financial costs of video understanding tasks
• →Active perception theory proves effective for sparse, temporally dispersed information common in real-world long video queries
• →Framework demonstrates efficiency gains suggest selective attention and agency outperform scale-only approaches in multimodal AI