WHY LANGUAGE MODELS HALLUCINATE

Summary from the TLDR AI newsletter

Hallucinations persist because current evaluation methods reward wild guessing that might occasionally be correct over admitting a lack of knowledge. The fix is simple: calibrate rewards during training based on certainty, penalize confidently wrong errors more, and give partial credit when models hedge or express doubt.