On September 4, 2026, researchers posted “Amortizing Scaling Law Construction Costs” (arXiv:2609.05016), a workshop paper proposing a cheaper way to derive the scaling laws labs use to plan model training runs. Instead of training a large, fixed grid of model-size and data-size combinations, the authors formulate data collection for scaling-law fitting as a Bayesian optimization problem.
The method combines two ideas. First, it progressively expands the compute budget during data acquisition, which mirrors how labs actually schedule training runs rather than assuming a uniform grid from the start. Second, it uses “surrogate-fantasized” evaluations that estimate the likely outcome of configurations the model has not actually trained, so the optimizer can pick the most informative next real training run instead of running an exhaustive sweep.
Against full-grid baselines, the approach recovers scaling-law fits that closely match the exhaustive method while requiring 10 to 100 times fewer training runs, without sacrificing the accuracy needed to guide architecture and data-mix decisions. It is a short, 5-page submission focused narrowly on the cost of the fitting process itself, not a new scaling law for any specific model family.
Scaling laws are how labs decide compute allocation before committing to a nine- or ten-figure training run, so a 10-100x cut in the cost of deriving one directly lowers the cost of getting that decision right. This is a workshop-scale validation of the method, not yet a demonstrated scaling law behind a shipped frontier model - the real test will be whether labs report using it on live training decisions.