Grad Detect: Gradient-Based Hallucination Detection in LLMs

Grad Detect is a novel gradient-based method for detecting hallucinations in Large Language Models (LLMs) by analyzing layer-wise gradient patterns during inference. It leverages the internal gradient structure, which reveals information about output correctness that is not available from output-level signals, and demonstrates superior performance over existing confidence-based and sampling-based methods on various Q&A benchmarks. The approach emphasizes the final five layers of the model, which contain over 97% of the relevant gradient signal, facilitating efficient implementation with minimal performance degradation, thus enhancing the reliability of LLMs in critical applications.