This research introduces a novel approach to detect token-level hallucinations in large language model outputs by analyzing patterns in attention matrices.

• Extracts features from attention matrices to identify irregular patterns associated with hallucinations
• Examines both the average attention each token receives and the diversity of attention distribution
• Provides a method to pinpoint exactly where in a response an LLM may be fabricating information
• Contributes to making AI systems more secure and trustworthy for critical applications

From a security perspective, this approach enables more reliable AI deployments by allowing systems to flag potentially false information before it reaches users, reducing risks in high-stakes domains like healthcare or financial services.

Hallucination Detection using Multi-View Attention Features