Anomaly Detection: Navigating the Edge of Normalcy Across Diverse Domains

Latest 50 papers on anomaly detection: Oct. 20, 2025

Anomaly detection is the bedrock of robust AI/ML systems, crucial for everything from safeguarding critical infrastructure to enabling scientific discovery. As data grows in volume, velocity, and complexity, distinguishing the normal from the truly anomalous becomes an ever-present, evolving challenge. Recent research offers exciting breakthroughs, pushing the boundaries of what’s possible, from leveraging advanced language models to pioneering novel geometric representations.

The Big Ideas & Core Innovations

At the heart of these advancements lies a common quest: to identify rare, unusual, or suspicious patterns in data more accurately, efficiently, and explainably. Several papers spotlight the power of multimodal and contextual understanding in surfacing complex anomalies. Researchers from the University of South Florida and Mitsubishi Electric Research Laboratories (MERL) in their work, “Leveraging Multimodal LLM Descriptions of Activity for Explainable Semi-Supervised Video Anomaly Detection”, demonstrate how Multimodal Large Language Models (MLLMs) can generate high-level, interpretable textual descriptions of object activities to detect complex interaction-based anomalies in videos. Similarly, “An LLM-Powered AI Agent Framework for Holistic IoT Traffic Interpretation” proposes an LLM-powered AI agent framework for IoT traffic, moving beyond traditional threat detection by providing deeper operational understanding through natural language processing. In a similar vein, “Semantic Visual Anomaly Detection and Reasoning in AI-Generated Images” introduces AnomAgent, a multi-agent framework that reasons about commonsense knowledge and physical feasibility to detect and explain semantic anomalies in AI-generated images, showcasing the shift towards explainable and context-aware anomaly detection.

Another significant theme is enhanced robustness and scalability through advanced modeling. The paper, “Isolation-based Spherical Ensemble Representations for Anomaly Detection”, from Tsinghua University and Great Bay University, introduces ISER, a method leveraging spherical ensemble representations and density-aware scoring to improve computational efficiency and handle diverse anomaly types with linear time complexity. For time series, “CrossAD: Time Series Anomaly Detection with Cross-scale Associations and Cross-window Modeling” by researchers from East China Normal University, offers a framework that explicitly models cross-scale and cross-window associations to capture dynamic relationships, overcoming limitations of fixed window sizes. In cybersecurity, the authors from the University of Calgary, in “Improving Intrusion Detection with Domain-Invariant Representation Learning in Latent Space”, demonstrate improved out-of-distribution generalization by learning domain-invariant representations.

Medical imaging sees a surge in innovative techniques for delicate anomaly detection. “Generating healthy counterfactuals with denoising diffusion bridge models” by researchers from Massachusetts General Hospital and Harvard Medical School, introduces DDBMs to generate healthy counterfactuals from pathological MRI data, significantly improving segmentation and anomaly detection. “Unsupervised Deep Generative Models for Anomaly Detection in Neuroimaging: A Systematic Scoping Review” reviews the rapid advancements of generative models in neuroimaging, particularly highlighting anatomy-aware designs and pseudo-healthy reconstruction. A fresh perspective is offered by Tsinghua University researchers in “Rethinking Medical Anomaly Detection in Brain MRI: An Image Quality Assessment Perspective”, which argues for integrating image quality assessment into anomaly detection for more reliable diagnostics.

Under the Hood: Models, Datasets, & Benchmarks

This collection of papers highlights a rich landscape of innovative models, specialized datasets, and crucial benchmarks that are propelling anomaly detection forward:

• MLLMs for Explainability: “Leveraging Multimodal LLM Descriptions of Activity for Explainable Semi-Supervised Video Anomaly Detection” utilizes MLLMs like Gemma 3 (outperforming GPT-4o for this task) to generate high-level textual descriptions of video events, enhancing interpretability.
• Cross-Modal Distillation (CMDIAD): In “Incomplete Multimodal Industrial Anomaly Detection via Cross-Modal Distillation”, Thermo Fisher Scientific and Technical University of Denmark researchers propose CMDIAD, a framework that trains with multimodal data (RGB and 3D point clouds) but infers with fewer modalities, crucial for industrial quality control. Code available at https://github.com/evenrose/CMDIAD.
• Hybrid Geometric GNNs (Janus): “Combining Euclidean and Hyperbolic Representations for Node-level Anomaly Detection” from University of Calabria and University of Udine introduces Janus, a Graph Autoencoder that combines Euclidean and hyperbolic latent spaces for node-level anomaly detection. Code can be found at https://anonymous.4open.science/r/JANUS-5EDF/.
• Reservoir Computing with Spectral Residuals (SR-RC): Chiba Institute of Technology researchers in “Enhancing Time-Series Anomaly Detection by Integrating Spectral-Residual Bottom-Up Attention with Reservoir Computing” propose SR-RC, integrating a learning-free spectral residual attention mechanism with reservoir computing for efficient time-series anomaly detection, suitable for edge AI.
• Neuromorphic Systems: “Online Reliable Anomaly Detection via Neuromorphic Sensing and Communications” explores neuromorphic computing for real-time anomaly detection, leveraging event-based data processing.
• Causal Digital Twins (CDT): University of Extremadura and University of Verona in “Causal Digital Twins for Cyber-Physical Security: A Framework for Robust Anomaly Detection in Industrial Control Systems” introduce CDT, a framework for industrial control systems that achieves high F1-scores and root cause analysis accuracy on datasets like SWaT.
• Federated eBPF Monitoring (FedMon): “FedMon: Federated eBPF Monitoring for Distributed Anomaly Detection in Multi-Cluster Cloud Environments” presents FedMon, a federated learning framework using eBPF for privacy-preserving distributed anomaly detection in multi-cluster clouds.
• Domain-Invariant Learning: “Improving Intrusion Detection with Domain-Invariant Representation Learning in Latent Space” uses mutual information regularization and reconstruction loss to create domain-invariant latent spaces for OOD intrusion detection. Code at https://github.com/padmaksha18/MTRAE/blob/main/mtrae/mtl-reg-cse-cic-ids.
• NLP Anomaly Benchmarks: “AD-LLM: Benchmarking Large Language Models for Anomaly Detection” and “NLP-ADBench: NLP Anomaly Detection Benchmark” introduce the first LLM-based and comprehensive NLP anomaly detection benchmarks, respectively, evaluating performance on tasks like zero-shot detection and data augmentation. The latter includes eight datasets and 19 algorithms, with code at https://github.com/USC-FORTIS/NLP-ADBench.
• Denoising Diffusion Bridge Models (DDBMs): “Generating healthy counterfactuals with denoising diffusion bridge models” uses DDBMs to generate healthy MRI counterfactuals, outperforming traditional diffusion models.
• RFOD for Tabular Data: National University of Singapore researchers in “RFOD: Random Forest-based Outlier Detection for Tabular Data” propose RFOD, a Random Forest framework for mixed-type tabular data, using feature-wise conditional reconstruction and uncertainty-weighted scoring.

Impact & The Road Ahead

This wave of research demonstrates a clear trajectory: anomaly detection is becoming more intelligent, robust, and domain-aware. The integration of advanced AI techniques, from multimodal LLMs to causal inference and hybrid geometric embeddings, is transforming how we detect the unusual. These advancements have profound implications across diverse fields: enhancing medical diagnostics, securing critical cyber-physical systems, improving industrial quality control, protecting autonomous vehicles from stealthy attacks, and even monitoring marine life from space. The open-source contributions from many of these papers, such as those from CMDIAD (https://github.com/evenrose/CMDIAD) and Whale detection (https://github.com/microsoft/whales), further accelerate innovation.

The road ahead involves further pushing the boundaries of interpretability, ensuring models can explain why something is anomalous, not just that it is. Addressing adversarial robustness, as highlighted by “Towards Adversarial Robustness and Uncertainty Quantification in DINOv2-based Few-Shot Anomaly Detection”, will be critical for safety-critical applications. Furthermore, the development of more diverse and challenging benchmarks, exemplified by ASBench (https://arxiv.org/pdf/2510.07927) for synthetic image anomalies and FetalSigma-1M (https://arxiv.org/pdf/2510.12953) for fetal ultrasound, will be essential to drive innovation. As AI continues to permeate every facet of our world, the ability to reliably identify and understand anomalies will remain a cornerstone of trusted and intelligent systems.