Language Models in Video Anomaly Detection: A Dynamic Survey

Abstract

Video Anomaly Detection (VAD) plays a critical role in intelligent surveillance and public safety by identifying rare, unexpected, or contextually abnormal events in complex video environments. Conventional deep learning approaches have substantially advanced VAD, but they remain largely driven by visual representations and often face challenges in semantic generalization, interpretability, and adaptation to unseen anomaly types. Recent Large Language Models (LLMs) and Multimodal Large Language Models (MLLMs) are reshaping the field by enabling semantic reasoning, contextual understanding, language-guided detection, open-vocabulary recognition, and natural-language explanation. This survey provides a focused and systematic review of language-model-based VAD. We organize existing methods across major supervision and deployment paradigms, including fully supervised, unsupervised, semi-supervised, weakly supervised, training-free, instruction-tuned, and open-world/open-vocabulary VAD, while positioning them relative to key non-language-model foundations. We further review representative datasets and evaluation protocols, emphasizing their suitability and limitations for both anomaly detection and anomaly understanding. Beyond serving as a static snapshot of the literature, this paper is designed as a dynamic survey whose taxonomy, method coverage, datasets, and evaluation discussion can be maintained through versioned updates as new work emerges. This evolving resource is supported by the public website dynamicvadsurvey.github.io, which provides access to maintained releases, archived versions, update records, and interactive exploration of the surveyed literature.