10). Hidden Factual Knowledge in LLMs This study introduces a framework to measure hidden knowledge in LLMs, showing that models encode significantly more factual information internally than they express in outputs, up to 40% more.

→ View original post on X — @dair_ai