Stanford AI Lab researchers led by Suze van Adrichem introduced hinted scaling laws that jointly model compute scale C and hint level H to forecast GSM8K accuracy within three points from models scoring zero using 5600 times less compute.