Language models understand causality internally but cannot say it aloud

Original: Causal Tongue-Tie: LLMs Can Encode Causal Direction, But Their Yes/No Outputs Fail to Express

Source: arxiv.org ↗