VideoARM: Agentic Reasoning over Hierarchical Memory for Long-Form Video Understanding

Status: Accepted to CVPR 2026
arXiv: https://arxiv.org/abs/2512.12360
Code: https://github.com/qiankemeng/VideoARM

Background

Long-form video understanding requires models to preserve key events across long temporal spans, connect information across segments, and perform multi-step reasoning when answering questions. Feeding many raw frames directly into context is costly and often leads to noise accumulation and missed key evidence.

VideoARM addresses this problem with a hierarchical memory and agentic reasoning framework for long-video question answering. The framework enables controlled retrieval and reasoning over compressed video memories instead of consuming the whole video context at once.

Core Ideas

• Hierarchical video memory: Organize long-video information into event-, segment-, and global-level semantic memories.
• Agentic reasoning: Let the agent dynamically access relevant memories according to the question.
• Long-video QA loop: Unify visual information extraction, memory organization, and reasoning-based answering in an extensible pipeline.

Current Role

This work is one of my main research outputs on long-form video understanding and multimodal agents. It also motivates my ongoing exploration of video memory, long-video multi-agent systems, and visualization-assisted evaluation.

---

Keywords: Long-Form Video Understanding | Agentic Reasoning | Hierarchical Memory | Video QA | CVPR 2026