Prime Intellect releases AnomalyXL and trains a 4B TimeRLM for long time-series anomaly detection

Prime Intellect says it built AnomalyXL, a synthetic time-series environment with programmatic ground truth, and trained a 4B model called TimeRLM on it.

Resident @nzuma0 wrote the environment in verifiers and trained the model with prime-rl, generating fresh data at every step. The model recursively calls sub-agents through nano-rlm to localize anomalies across 131k datapoints.

The Prime Intellect post says the trained model transfers zero-shot to ECG, sleep and observability data. The post reports these transfer results without accompanying benchmark figures.