Khái niệm cốt lõi
GENAUDIT is a tool designed to assist in fact-checking LLM responses by identifying errors and providing evidence to support or refute claims.
Thống kê
LLMs can generate factually incorrect statements even with access to reference documents.
Errors detected by GENAUDIT: 8 different LLM outputs.
GENAUDIT highlighted ∼40% of erroneous words with ∼95% precision.
Achieved ∼91% recall and ∼95% precision in extracting useful evidence.
Trích dẫn
"Such errors can be dangerous in high-stakes applications."
"We release our tool (GENAUDIT) and fact-checking model for public use."