ReactVAU: A Slow-Fast Decoupled Framework for Streaming Video Anomaly Understanding
Chia-Hui Chen, Shih-Ying Yeh, Fu-En Yang, Min-Hung Chen, Shang-Hong Lai
National Tsing Hua University · NVIDIA
Published September 7, 2026 · on the feed September 9, 2026 · ▲ 17 · arXiv 2609.07941
The abstract
In this paper, we propose ReactVAU, a Slow-Fast Decoupled Framework for real-time streaming Video Anomaly Understanding (VAU). Existing VAU methods rely on offline inference with global temporal sampling, which violates causality and prevents deployment in live surveillance streams. Conversely, general streaming video models satisfy causal access but dilute rare transient anomalies during memory compression and often invoke heavyweight MLLMs uniformly over long normal intervals. React VAU addresses this gap with three synergistic components: a lightweight Fast Detection Module based on Spatial Grid Folding (SGF) for continuous anomaly filtering; an Anomaly-Aware Persistent Memory (AAPM) that protects critical visual cues from temporal decay; and a heavyweight Slow Reasoning Module that remains dormant during normal streams and is awakened only by suspicious events for semantic verification and causal description. Extensive experiments on multiple benchmarks demonstrate that ReactVAU operates under strict streaming constraints while simultaneously achieving competitive performance in both anomaly detection and causal reasoning, alongside significantly enhanced computational efficiency by minimizing heavyweight MLLM invocations. Project page is available at https://huiyuiui.github.io/React_VAU/
The Brief
2:19A short podcast summary of the paper. Written and voiced by AI.
Transcript
ReactVAU. A Slow-Fast Decoupled Framework for Streaming Video Anomaly Understanding. By Chia-Hui Chen and four others, from National Tsing Hua University and NVIDIA. Published on the seventh of September. Its code repository is worked out: about sixty source files, with a readme and a licence.
Video Anomaly Understanding, or VAU, means detecting unusual events in video and explaining them. Existing VAU methods run offline, sampling globally across the whole video. That breaks causality, because the method looks at frames that have not happened yet, and it rules out deployment on live surveillance streams. General streaming video models respect causal access, but they dilute rare, transient anomalies when they compress memory, and they often call heavyweight MLLMs uniformly, even across long intervals where nothing abnormal happens. One family cannot run live, and the other can run live but misses what matters and spends heavily while doing so.
ReactVAU is a slow-fast decoupled framework for real-time streaming anomaly understanding. A cheap fast path watches everything, and an expensive slow path is used only when something looks wrong. Detection is continuous and cheap, and deeper reasoning is reserved for the events that call for it.
It has three components that work together. A lightweight Fast Detection Module, based on Spatial Grid Folding, or SGF, filters continuously for anomalies. An Anomaly-Aware Persistent Memory, or AAPM, protects critical visual cues from fading as time passes. And a heavyweight Slow Reasoning Module stays dormant during normal stretches of the stream. It is woken only by suspicious events, and then carries out semantic verification and describes what caused the event.
Extensive experiments on multiple benchmarks show ReactVAU working under strict streaming constraints while achieving competitive performance in both anomaly detection and causal reasoning. It also significantly improves computational efficiency, by keeping heavyweight MLLM invocations to a minimum.
The division of labour answers both earlier failures. The persistent memory keeps rare anomalies from being washed out, which was the streaming models' weakness. The dormant reasoner saves the costly analysis for the moments that need it, instead of spending it on hours of normal footage. And the whole design only ever looks backward in time, which is what live deployment requires.
Paper Quest · 7 chapters · about 41 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
Hours of empty corridor, then a few seconds that matter. Your system has to catch them live.
You watch live surveillance streams and want a system that flags unusual events and explains what caused them. The best explainers run offline, sampling the whole video, including frames that have not happened yet.
The road ahead
- 1 · The problem ~7 min Live surveillance needs a system that flags and explains unusual events as they happen, but today's tools either peek at the future or waste effort and forget brief events.
- 2 · What came before ~3 min
- 3 · The key idea ~9 min A cheap detector watches every frame, a protective memory keeps suspicious evidence, and a heavy reasoner wakes only for flagged events to verify and explain them.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Each second the fast detector scores a frame grid, memory guards suspicious cues, a high score wakes the slow reasoner, and it verifies and describes the event.
- 6 · The experiment ~6 min You’ll guess: Working live without any future frames, how did ReactVAU's detection and reasoning compare with existing methods, and at what cost?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.