A new research framework skips the hunt for one universal anomaly detector and instead builds a dispatcher that picks the right specialist for the job.
The framework, called TS-Router, is described in an arXiv paper from researchers in cs.AI. Instead of using a single scoring method to flag anomalies across every dataset, it uses a pretrained time-series foundation model to read a dataset's underlying patterns, then estimates which of several specialist anomaly detectors is best suited to that specific data. The catch is that there are no real anomaly labels to train this routing decision on, so the team instead generated labeled simulated tasks and used the specialists' relative performance on those to teach the router which detector tends to win in which situation. Once deployed, the router needs no target labels at all: it picks a shortlist of specialists and fits them to the new series unsupervised. Tested across 16 real-world benchmarks and four evaluation metrics, TS-Router came out with the best average rank among the approaches compared.
This matters because time-series foundation models have mostly chased one-size-fits-all scoring, the same instinct that drives most foundation-model hype. Anomaly detection does not cooperate with that instinct: a spike that is normal in server CPU metrics is a meltdown in a patient's heart-rate sensor. Treating detector selection as the thing to learn, rather than detection itself, is a more honest match for how messy real-world time series actually behave.
Still, "best average rank on 16 benchmarks" is an academic scoreboard, not a production guarantee, and the code sits in an anonymous repository rather than under a named lab or company. Whether this routing idea survives contact with genuinely chaotic industrial data, instead of curated benchmark sets, is the test that actually matters.