LAKE: Training-Free Anomaly Detection in Vision-Language Models

A recent study on arXiv disputes the belief that vision-language models (VLMs) need additional adapters or memory systems for anomaly detection (AD). The researchers contend that pre-trained models inherently contain anomaly-specific knowledge, which is often dormant and not fully utilized, residing in a limited group of neurons that respond to anomalies. They introduce a framework called latent anomaly knowledge excavation (LAKE), which operates without training and utilizes a small number of normal samples to activate these essential neuronal signals. By targeting these responsive neurons, LAKE creates a highly efficient representation of normality. This research is documented in arXiv:2604.07802v3.