Researchers from Google DeepMind and Stanford University have introduced the Search-Augmented Factuality Evaluator (SAFE) framework, which automates the evaluation process to assess the factuality of content generated by large language models. Using Google Search as a reference point, SAFE breaks down long-form responses into individual facts and verifies their accuracy. The framework aligns with human assessments in evaluating factual information and offers a scalable and cost-efficient method for factuality evaluation.
