#evaluation-incentives

#evaluation-incentives

Large language models produce confident but false statements because pretraining emphasizes next-word prediction and current evaluation incentives reward superficially correct outputs.