Russell Bent

LLMs can iteratively refine their outputs to reduce hallucinations by optimizing for semantic consistency with trusted contexts, achieving better reliability than likelihood-based or retrieval-augmented methods.