Xining Xun

Internal representations in large language models can be read with high accuracy long before they become causally effective, revealing a critical disconnect in model interpretability.