ReactVAU at ECCV: slow-fast decoupling cuts MLLM calls for streaming video anomaly understanding
CMHungSteven · x · 2026-09-10
ReactVAU, an ECCV paper from National Tsing Hua University and NVIDIA, introduces a slow-fast decoupled framework for causal streaming Video Anomaly Understanding:
- Problem: conventional VAU needs the full video and wakes a heavy MLLM on every normal second, conflicting with real deployment where future frames are unavailable.
- Architecture: a Fast module built on PaliGemma2-3B uses Spatial Grid Folding (four frames at 4 FPS folded into a 2x2 grid) to score every second; persistent memory stores rare cues; the slow reasoning module fires only when the stream looks suspicious.
- Results: follows a causal protocol, achieves competitive detection and descriptions while drastically reducing MLLM calls.
- Code and checkpoints are open-sourced on GitHub and Hugging Face.
More from Research
- Token Collapse: Study Finds Graph Soft Prompting Tokens in LLMs Fail to Encode Structure — Sapienza · 2026-10-27
- Terence Tao's SAIR Foundation revamps Mathathon with $1.5M compute per team — _sathvikr · 2026-10-27
- Mathathon 2.0 launches with Terence Tao's SAIR, $1.5M in compute for teams — _sathvikr · 2026-10-27
- The 'sweet lesson' is real: two kinds of ML problems need opposite strategies — mike64_t · 2026-10-27
- PixGR: Pixel-space generative decoding for fast feed-forward 3D reconstruction — kwangmoo_yi · 2026-10-27
- UniAction: unified action conditioning lets one world model handle mixed action types — ZeYanjie · 2026-10-27