Task Priors: Enhancing Model Evaluation by Considering the Entire Space of Downstream Tasks
This research introduces Task Priors, a probabilistic framework that replaces fixed benchmarks with continuous task distributions to rigorously evaluate AI m...