arXiv Notify me• Jun 17, 2026paper

Native Active Perception as Reasoning for Omni-Modal Understanding

Executive Summary

Continue your research

Keep pulling the thread on Junyang Lin.

5 quotes

Enabling efficient and scalable analysis of long-form video content, a major bottleneck for current models.
Pioneering a new class of smaller, more efficient agentic models that can outperform larger, passive models.
Applying the active perception framework to other data modalities beyond video, such as large documents or complex codebases.
Developing systems that can dynamically adjust computational spend based on query difficulty due to positive test-time scaling.

OmniAgent achieves state-of-the-art performance among open-source models across ten benchmarks, including VideoMME and LVBench.

The 7B parameter version of OmniAgent scored 50.5% on the LVBench benchmark, outperforming the 72B parameter Qwen2.5-VL-72B model's score of 47.3%.

OmniAgent is a native omni-modal agent that formulates video understanding as a POMDP-based iterative Observation-Thought-Action cycle.

OmniAgent uses on-demand actions to distill audio-visual cues into a textual memory, decoupling its reasoning complexity from video duration.

OmniAgent exhibits positive test-time scaling, where its performance improves as the number of reasoning turns increases.

Processed Jul 28, 2026Daily intelligence brief → yt-dlp + mlx-whisper + Gemini