The paper introduces a training-free approach to language-based video anomaly detection (VAD). Instead of fine-tuning model weights for a specific anomaly-detection task, it uses frozen vision-language models (VLMs) and large language models (LLMs) to localize abnormal events. The source explicitly frames this as avoiding task-specific model-weight optimization.
The title points to two components—anomaly-guided observation and two-level semantic feedback—but the abstract cuts off before describing them. The only concrete claims available are the training-free setup and the reliance on frozen models; no results, comparisons, or implementation details are included in the source text.
Because there is only one source and its abstract is incomplete, this digest cannot assess how the method performs against existing VAD approaches. The significance lies in the direction: pushing language-based anomaly detection toward zero-optimization pipelines built on general-purpose pretrained models.