EventVAD introduces a training-free framework that uses large language models to detect anomalies in video footage without requiring domain-specific training data.

• Combines world knowledge from LLMs with fine-grained visual analysis to identify suspicious activities
• Focuses on event-aware detection that can recognize context-specific anomalies across diverse scenarios
• Overcomes limitations of both supervised approaches (which struggle with unseen anomalies) and existing LLM methods (which lack precise localization)
• Demonstrates superior generalization capabilities for security and surveillance applications

This research enables more flexible, adaptable security monitoring systems that can identify threats without extensive training datasets—critical for real-world deployments where anomalies are rare and diverse.

EventVAD: Training-Free Event-Aware Video Anomaly Detection