Lowest Span Confidence: Zero-Shot Hallucination Detection from a Single LLM Response
Lowest Span Confidence (LSC) is a new zero-shot method to detect LLM hallucinations using only one forward pass and output probabilities. It works by calculating the average confidence of short text segments within sliding windows of the model's output. LSC outperforms existing methods by capturing localized uncertainty patterns, making it practical for API use where internal model details are hidden.