@lamsheeper
AI
CATT's researchers find that averaging token scores can bury the few tokens that matter to an answer.