Bayesian Optimization for Efficient Scaling Law Construction

A new framework reduces the computational cost of deriving scaling laws by formulating data collection as a Bayesian optimization problem. By using surrogate-fantasized evaluations and progressive compute expansion, researchers can recover accurate scaling laws with fewer training runs.