@victorwang37
PhD student @UTAustin
AI
A research team reports that two confidence-estimation methods struggled on questions one model checkpoint answered correctly and another missed.