LLMDP-U turns LLM reasoning into reusable anomaly detectors

LLMDP-U uses a large language model as a compiler for time-series anomaly detection rather than keeping it involved in routine scoring. Given a statistical profile of an initial reference segment and a numerical interface, the system generates and validates a series-specific Python program. That program produces multiple continuous anomaly scores without further LLM calls, and a fixed, label-free rank-fusion method combines them into a final score.

In offline, transductive evaluation on TSB-AD-U Eval-350, Opus-compiled programs achieved 0.5498 Average VUS-PR, a +0.0500 gain over a no-LLM Canonical Program and the third-highest leaderboard result. Performance improved from Haiku to Sonnet to Opus under a shared protocol. A matched-size exhaustive bank of numerical primitives did not reproduce the Opus gain.

An audit of 1,194 programs identified recurring profile-conditioned mechanisms involving activity, persistence, and value support. The approach offers an alternative to both fixed numerical detectors and systems requiring recurring foundation-model inference: adaptation is captured in a persistent, inspectable program, while subsequent scoring remains numerical.