🔍 Anomaly Detection¶

🔬 ICLR2026 · 10 paper notes

📌 Same area in other venues: 📷 CVPR2026 (7)

🔥 Top topics: Anomaly Detection ×7 · Time-Series Forecasting ×2 · Few-/Zero-Shot Learning ×2

Adaptive Conformal Anomaly Detection with Time Series Foundation Models for Signal Monitoring: The authors propose W1-ACAS: a post-hoc, tuning-free adaptive conformal anomaly detection framework. It maps prediction errors from pre-trained Time Series Foundation Models (TSFMs) into anomaly scores directly interpretable as false positive rates (p-values) and learns weights online by minimizing the Wasserstein distance to maintain stable false positive control under non-stationary data.
Foundation Visual Encoders Are Secretly Few-Shot Anomaly Detectors: The authors discovered that frozen foundation visual encoders "secretly" possess the ability to distinguish anomalies—the area of an anomalous region in an image is positively correlated with the distance of its features to the natural image manifold. By training a lightweight non-linear projection operator (FOUNDAD) atop the encoder to pull anomalous features back to the normal manifold and scoring based on the difference before and after projection, SOTA performance is achieved in few-shot, category-agnostic industrial anomaly detection.
Healthcare Insurance Fraud Detection via Continual Fiedler Vector Graph Model: ConFVG utilizes the second smallest eigenvector of the graph Laplacian (Fiedler vector) to guide the masking strategy of a Graph Autoencoder (GAE) for structural-aware representation learning under label scarcity. It then employs subgraph attention fusion and a Mean Teacher framework to continuously adapt to evolving fraud patterns in unlabeled online streams, achieving real-time healthcare fraud detection.
Let OOD Feature Exploring Vast Predefined Classifiers: This paper proposes VPC, which utilizes a set of fixed equiangular prototypes to map ID classes and OOD samples into two distinct predefined subspaces. By using the difference in \(L_2\) activation intensity between these two subspaces as an OOD score, it consistently reduces FPR95 in Outlier Exposure (OE) training scenarios on CIFAR and ImageNet-1k.
LLM as an Algorithmist: Enhancing Anomaly Detectors via Programmatic Synthesis: This work repositions the LLM from a "data processor" to an "algorithmic strategist"—it analyzes the algorithmic description of a detector without touching real data, reasons about its logical blind spots, and generates a reusable Python synthesis code. This code creates "hard anomalies" specifically designed to deceive that detector, upgrading the original one-class problem into a more separable two-class problem. It consistently enhances five mainstream detectors across 36 tabular anomaly detection benchmarks.
Low Rank Transformer for Multivariate Time Series Anomaly Detection and Localization: This paper theoretically maps the learning process of Transformer encoders on multivariate time series to the classical STAR statistical model. It proposes ALoRa-T, which applies low-rank regularization to self-attention, using the "rank" of the attention matrix as an anomaly signal for detection and tracing anomalies back to specific variables for localization using interpretable contribution weights.
MRAD: Zero-Shot Anomaly Detection with Memory-Driven Retrieval: MRAD replaces the parametric fitting of \(p(y|x)\) in mainstream ZSAD with "similarity retrieval from a feature-label memory bank." The training-free version suppresses WinCLIP, and combined with two linear fine-tuning layers and dynamic prompts injected with regional priors, it achieves SOTA across 16 industrial and medical datasets.
PIRN: Prototypical-based Intra-modal Reconstruction with Normality Communication for Multi-modal Anomaly Detection.: PIRN targets few-shot multimodal industrial anomaly detection for RGB images and 3D surface normals. It reconstructs normal features of each modality using adaptive prototype codebooks and enhances texture and geometric cues through cross-modal normality communication, achieving superior detection and localization performance on MVTec 3D-AD, Eyecandies, and Real-IAD D3.
ReTabAD: A Benchmark for Restoring Semantic Context in Tabular Anomaly Detection: ReTabAD is the first "context-aware" tabular anomaly detection benchmark. It restores discarded textual semantics (feature descriptions, domain knowledge, original categorical values) into 20 curated datasets, provides implementations for 20 classic/deep/LLM-based algorithms, and proposes a training-free zero-shot LLM framework. Experiments demonstrate that semantic context improves detection AUROC by an average of 7.6 percentage points, allowing zero-shot LLMs to approach Prev. SOTAs.
UniOD: A Universal Model for Outlier Detection across Diverse Domains: UniOD trains one universal outlier detection model using a batch of historical labeled datasets. It first unifies tabular datasets of any dimension or semantics into "multi-scale similarity graphs + SVD features," then transforms outlier detection into node binary classification using a GIN+GT dual-path graph network. Once trained, the model performs training-free and parameter-tuning-free inference for any unseen new dataset, achieving average AUROC/AUPRC scores that outperform 17 baselines across 30 benchmarks with lower latency.