out-of-distribution detection

Amid the rise of vision-language models (VLMs/LVLMs), out-of-distribution (OOD) detection and related tasks—such as anomaly detection, novelty detection, open-set recognition, and outlier detection—suffer from conceptual ambiguity and paradigmatic fragmentation. This paper proposes “Generalized OOD Detection v2”, a unified framework that systematically clarifies semantic boundaries and evolutionary relationships among these tasks, identifies OOD detection and anomaly detection as the central challenges, and formalizes novel evaluation paradigms and problem settings introduced by LVLMs (e.g., GPT-4V). Leveraging CLIP-based semantic alignment analysis, task taxonomy modeling, benchmark evolution comparison, and cross-task method review, the work synthesizes over 100 studies. The resulting VLM-driven conceptual framework redefines foundational assumptions, pinpoints critical technical challenges—including semantic misalignment, evaluation inconsistency, and LVLM-specific failure modes—and charts concrete directions for future research, establishing itself as the definitive survey in this rapidly evolving domain.

This work addresses the challenge of unreliable sensor readings in industrial inspection robots caused by occlusions, limited viewpoints, or environmental anomalies, which hinder real-time task status assessment. The authors propose a hybrid framework that integrates supervised fault classification with unsupervised anomaly detection, uniquely combining conformal prediction and world models to enable policy-agnostic, distribution-free early discrimination among three states—success, known faults, and out-of-distribution anomalies—using compressed video inputs. The approach facilitates training data quality evaluation and model feedback, achieving over 90% recognition accuracy on both office and industrial instrument inspection datasets. It outperforms human observers in decision speed and has been successfully deployed on a Boston Dynamics Spot robot for real-time operation.

Current mainstream out-of-distribution (OOD) detection methods suffer from a fundamental objective misalignment: supervised models trained exclusively on in-distribution (ID) data erroneously equate high predictive uncertainty or large feature-space distances with OOD samples, leading to irreducible detection errors. Method: The authors systematically analyze major paradigms—including uncertainty estimation, feature-logit hybrid models, density estimation, and generative approaches—and rigorously prove their theoretical limitations under common distribution shifts. Contribution/Results: They introduce the “objective misalignment” framework, establishing that OOD detection is inherently a distribution discrimination task—not an uncertainty or distance regression problem. Crucially, they demonstrate that prevalent mitigation strategies—such as anomaly exposure, model architecture expansion, or cognitive uncertainty modeling—cannot rectify this foundational flaw. This work provides critical theoretical grounding for paradigmatic reformulation of OOD detection.

Existing out-of-distribution (OOD) research predominantly focuses on extrapolative (“outside”) anomalies while overlooking interpolative (“inside”) in-distribution anomalies—i.e., samples that reside within the support of the training distribution yet are semantically anomalous. Method: This work formally defines and empirically validates the “inside OOD” concept, proposing a two-dimensional analytical framework to distinguish inside from outside OOD. Leveraging statistical distribution analysis, geometric modeling in feature space, and multi-model robustness evaluation, we systematically characterize their co-occurrence patterns and differential impacts on model behavior. Results: We demonstrate that inside OOD triggers latent, progressive performance degradation, whereas outside OOD induces sharp confidence collapse. Our framework bridges a critical gap in conventional OOD detection, providing both theoretical grounding and empirical evidence for designing targeted defense mechanisms against distinct OOD categories.

Industrial anomaly detection is often hindered by the extreme scarcity of fault samples, leading to suboptimal model performance and limited generalization. This work addresses this challenge by constructing a problem-agnostic hyperspherical synthetic dataset to systematically evaluate 14 anomaly detection algorithms—including kNN, LOF, XGBOD, SVM, and CatBoost—under rigorously controlled conditions across varying fault rates (0.05%–20%) and training set sizes. The study quantitatively demonstrates for the first time that unsupervised methods achieve optimal performance when fewer than 20 fault samples are available; semi-supervised and supervised approaches significantly outperform others with 30–50 fault samples; and further increasing normal samples yields diminishing returns. Additionally, feature dimensionality is found to critically influence the efficacy of semi-supervised methods, thereby clarifying the operational boundaries of each algorithmic category.

This work addresses the performance gap between academic benchmarks and real-world deployment in unsupervised anomaly detection, where existing methods often exhibit instability, sensitivity to preprocessing, and inconsistent behavior in industrial settings. The authors conduct a systematic evaluation of 19 models on BowTie, a complex manufacturing dataset, revealing significant discrepancies between benchmark results and practical efficacy. To bridge this gap, they propose a human-in-the-loop unified detection framework that integrates SAM-generated refined candidate regions, heatmap-guided inspection, mask-based evaluation, and interactive verification. This framework enables quality inspectors to efficiently confirm defects, refine boundaries, and trace historical cases. Preliminary deployment demonstrates that the system substantially enhances both reliability and efficiency in industrial visual inspection.

Risk valuation systems are susceptible to undetected errors caused by data failures, misconfigurations, or anomalies, potentially leading to significant operational losses. This work proposes EQAF, a hierarchical unsupervised ensemble framework for anomaly detection that uniquely integrates domain-specific deterministic rules with multiple complementary statistical outlier detection methods to enable real-time integrity monitoring of risk computation outputs. EQAF effectively identifies subtle anomalies—such as “frozen values”—that are often missed by conventional purely statistical approaches. Experimental evaluation on four real-world risk datasets demonstrates that EQAF achieves F1 scores between 61% and 79% and improves AUC-ROC by 4–6 percentage points over the best individual baseline method, substantiating its robustness and effectiveness.

Current evaluation practices for supervised learning models are often misleading due to an overreliance on single aggregate metrics, which neglect the alignment among data characteristics, task objectives, and real-world application contexts. This work reframes model evaluation as a context-dependent, decision-oriented process and systematically investigates—through controlled experiments—the impact of dataset properties, validation strategies, class imbalance, and asymmetric error costs on evaluation outcomes. Leveraging diverse benchmark datasets, multiple validation protocols, and multidimensional performance measures, the study uncovers common pitfalls such as the accuracy paradox, data leakage, and metric misuse. It proposes a structured evaluation framework explicitly aligned with operational goals, offering principled guidance for developing more robust, reliable, and trustworthy supervised learning systems.

This work proposes a novel paradigm for anomaly detection, termed On-Model AD, which leverages the intrinsic knowledge of the primary model—specifically, the normal output ranges of its neurons—to detect anomalies without requiring a separate, dedicated detection model. Traditional approaches often deploy independent anomaly detectors, overlooking the rich distributional information already embedded within the main model, thereby introducing redundancy and inefficiency. In contrast, the proposed method eliminates the need for additional training or deployment overhead. Building upon this paradigm, we introduce RangeAD, an algorithm that achieves strong detection performance in high-dimensional tasks while substantially reducing inference costs, effectively balancing accuracy and computational efficiency.