Stanford AI Lab: Hinted Scaling Laws Improve GSM8K Forecasts
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.
SourceAnalysis
Stanford AI Lab researchers Suze van Adrichem, Luke Bailey and Rahul Pudi converted standard benchmarks into easier hinted variants and jointly modeled compute scale C with hint level H, enabling GSM8K accuracy forecasts at full difficulty within three points from 5600 times less training compute.
Stanford AI Lab
@StanfordAILabThe Stanford Artificial Intelligence Laboratory (SAIL), a leading #AI lab since 1963.