arXivLLMs & Text

Prompt-Response Semantic Divergence Metrics for Faithfulness Hallucination and Misalignment Detection in Large Language Models

The paper presents Semantic Divergence Metrics (SDM), a new system that improves the detection of significant deviations in Large Language Models' responses from the input context by measuring consistency across various semantically equivalent paraphrases.