Advances in Large Model Hallucination Research: New Techniques Reduce Hallucination Rate by 60%

Stanford University and Anthropic jointly released research on hallucination mitigation, combining retrieval augmentation, self-consistency checks, and uncertainty quantification to reduce the hallucination rate on the TruthfulQA benchmark from 35% to 14%. The study found that hallucinations mainly occur near the model's "knowledge boundary"; actively expressing uncertainty is a trainable capability; and RAG combined with citation tracking yields the best results.