#Good–Turing

'OpenAI research shows hallucinations are statistical consequences of pretraining and that binary benchmarking incentivizes overconfident guessing; changing evaluation to reward calibrated uncertainty can reduce hallucinations.'