Faithfulness Metrics for Hallucination Assessment in Large Language Models
A survey of how the faithfulness of LLM-generated text (its freedom from hallucination relative to a source) is measured across summarization, question-answering and machine translation. Covers lexical-overlap metrics (ROUGE, BLEU), embedding-based metrics (BERTScore), claim-extraction and QA-based metrics, and using an LLM itself as an evaluator, comparing each against human judgement, and reviews hallucination-mitigation strategies such as retrieval-augmented generation and prompting frameworks.
2501.00269
This review surveys how faithfulness (the absence of hallucination relative to source content) is evaluated for Large Language Model outputs across three generation tasks: summarization, question-ans…