OmniAgent Introduces Native Active Perception for Enhanced Video Understanding

The research presents OmniAgent, a pioneering omni-modal agent designed for long video understanding, which utilizes a POMDP-based iterative Observation-Thought-Action cycle. This approach allows for selective distillation of audio-visual cues into a persistent textual memory, decoupling reasoning complexity from video duration. Key innovations include Agentic Supervised Fine-Tuning for active perception and Agentic Reinforcement Learning with TAURA for optimized credit assignment. Empirical results across ten benchmarks demonstrate that OmniAgent achieves state-of-the-art performance, outperforming larger models such as Qwen2.5-VL-72B on specific tasks, including a notable 50.5% accuracy on LVBench.