The Model Agreed, But Didn’t Learn: Diagnosing Surface Compliance in Large Language Models
arXiv:2604.05995v1 Announce Type: cross
Abstract: Large Language Models (LLMs) internalize vast world knowledge as parametric memory, yet inevitably inherit the staleness and errors of their source corpora. Consequently, ensuring the reliability and m…