ReactVAU at ECCV 2026: fast-slow streaming video understanding cuts heavy MLLM calls
CMHungSteven · x · 2026-09-10
Presented at ECCV 2026 in Malmö, ReactVAU is a causal framework for streaming video understanding that avoids peeking at the future and avoids waking a heavy MLLM every normal second.
Design:
- An always-on fast detector monitors every frame;
- Persistent memory captures rare cues;
- Slow LLM-based reasoning fires only when the stream looks suspicious.
Under a causal protocol it achieves competitive detection and description performance with far fewer MLLM calls. Paper, code, weights, and demo are public; a joint NVIDIA × NTHU team effort, poster Thursday Sep 10, 16:30–18:30 CEST.
More from Research
- Stanford's PaperFormalizer turns research papers into Lean-verified versions with author in the loop — StanfordUniversity · 2026-10-27
- JockeyBench uses Japanese horse racing to test LLM agents — all betting strategies lose money — Zhongwei Xie · 2026-10-27
- Cohere's Invent-A-Dataset paper: zero-seed synthetic data beats frontier APIs on quality and diversity — idanbeck · 2026-10-27
- Hugging Face TRL makes score centering + masked importance sampling default in async GRPO — _lewtun · 2026-10-27
- AutoResearch agent's weird training curve gets dug into instead of discarded — ZimingLiu11 · 2026-10-27
- EMNLP 2026: sub-billion-parameter LMs show superhuman entity tracking, human-like fleeting memory — GretaTuckute · 2026-10-27