ReactVAU wakes its heaviest model only when a video stream turns suspicious

Offline sampling and memory compression both fail a live feed

A surveillance camera pointed at an empty hallway for six hours produces six hours of nothing. A video anomaly system that runs a multimodal large language model over all of it spends nearly its entire inference budget describing that nothing. ReactVAU, a newly posted paper on streaming video anomaly understanding, argues the fix is a scheduling question: the expensive reasoning should not run until something worth reasoning about appears.

Video anomaly understanding asks a system to detect an unusual event and explain it. That is a heavier job than flagging a timestamp. The ReactVAU abstract frames the field as caught between two designs that each break in a live setting. Offline methods sample a clip globally and reason over it all at once, which the authors say violates causality and blocks deployment in live streams. Streaming video models keep the causal order, but they dilute rare, short-lived anomalies as they compress their memory, and they tend to fire a heavy model on a fixed schedule whether or not anything is happening. The same argument shows up in work that splits long-term memory from reasoning: a cheap model fails less because it is small than because the memory feeding it was never built to hold context that lasts.

ReactVAU splits those two jobs into a fast path that never stops and a slow path that mostly sleeps.

Fast filter, slow reasoner, persistent memory

The first piece is a lightweight Fast Detection Module built on Spatial Grid Folding (SGF). It runs continuously, filtering the stream for anything that looks off without invoking a large model. The third piece is the heavyweight Slow Reasoning Module, which does semantic verification and causal description. That is the part that explains what was seen, and it stays dormant through normal footage, waking only when the fast module flags a suspicious event. Keeping the always-on half small is the whole point, and it is the same bet behind the current rush of sub-200M parameter models.

Between the two sits the Anomaly-Aware Persistent Memory, or AAPM. Its job is to keep the visual details that matter from fading as the stream compresses its own history. That is exactly where the streaming models the paper criticizes lose a transient event: by the time a heavy model would look, the cue has already been averaged into the background. AAPM is the paper's attempt to hold on to it, and the failure it is designed against is the one described in work on memory that stays retrievable after it goes stale.

The goal is a familiar one in production AI. Running a large model costs money, and running it on data that does not need it is the simplest way to waste that money. Routing systems are already built on the same instinct: compute-balanced routing matched best-of-16 voting while spending 58.9% fewer tokens.

Live surveillance is an unusual target for this kind of budget math. The feed never stops, the interesting moments are rare and brief, and a system that only explains events well after they happened is of little use to whoever is watching. ReactVAU's bet is that the rare event should be what triggers the expensive computation, and that everything else can be handled by something small.

The efficiency claim comes without a number

The paper states that experiments across "multiple benchmarks" show ReactVAU operating under streaming constraints while staying competitive on anomaly detection and causal reasoning, with "significantly enhanced computational efficiency" from invoking the heavy model less often. The abstract names no figures. There is no frames-per-second count, no accuracy number, no latency, and no list of which benchmarks the authors used.

That is normal for an abstract, but it leaves the efficiency claim as a promise rather than a measurement a reader can check. The project page at huiyuiui.github.io/React_VAU/ is where supporting material would live.

The architecture also concentrates its risk in the cheap half. If the fast module misses a short anomaly, the slow module never wakes and the system reports a normal stream. How often the filter misses, and how the memory holds up over hours rather than minutes, are the questions that decide whether a dormant-model design is a genuine saving or a delayed bill. The tension is not unique to video. Reasoning on a six-token budget buys latency and costs precision, and no architecture has yet escaped the trade. None of that is settled by an abstract that promises efficiency without a number attached to it.