A revised arXiv paper proposes a graph neural network framework for diagnosing which specific sensor or component is responsible for anomalies in multivariate industrial time series. The work aims to go beyond simple anomaly detection by making the source of the fault interpretable. It targets reliability and safety monitoring in complex, multi-sensor process environments.
A new arXiv preprint proposes combining classical template matching with variational techniques and statistical fusion to detect anomalies in patterned images. The authors argue that deep learning is often too costly or impractical when training data is scarce, while traditional template matching is interpretable but brittle to changes in scale and geometry. The method aims to keep the simplicity of template-based approaches while improving robustness to such variations.
Researchers propose a Lagrangian sub-flow framework built on continuous normalizing flows to detect out-of-distribution observations that lie in a subspace of high-dimensional data. The approach applies local diagnostics to the flow, aiming to better separate in-distribution samples from anomalous ones. The work is presented as a preprint on arXiv.
A revised arXiv preprint (2510.13311v2) introduces an unsupervised approach to detecting anomalies in tabular data by combining isolation-based techniques with spherical ensemble representations. The authors argue that existing unsupervised detectors still face fundamental limitations, and position their method for use cases such as offensive language detection, network security, and quality control. The work is a research contribution rather than a released product.
A new arXiv paper introduces Skynet, a method that detects failures in agentic AI systems by modeling both the semantics and the structure of multi-step workflows. Rather than judging individual outputs, it treats long-horizon plans, tool calls, and multi-agent coordination as a whole, since a single bad step such as an injected prompt or a flawed plan can derail the entire task. The authors position workflow-level monitoring as a way to catch these faults before they propagate.
Researchers have introduced an unsupervised framework for detecting anomalies in spacecraft telemetry that does not rely on labeled historical anomalies or lengthy warm-up periods, addressing common barriers to real-world deployment. The approach uses structure-aware modeling combined with adaptive Extreme Value Theory (EVT) thresholding to determine when telemetry readings should be flagged. The authors position the method as ready for operational use in settings where annotated failure data is scarce.
A new arXiv preprint presents PeriodicCALM, a framework for detecting anomalies on the fly in cyclostationary data streams, whose statistical properties vary periodically over time. Rather than relying on classical stationary assumptions, the algorithm adapts to these recurring temporal patterns as it monitors live data for deviations. The work was posted to arXiv's machine learning listing as a cross-listed submission.