Preprint reports lower-compute method for LLM scaling laws

A preprint describes Power-Law Entropy Search, a multi-fidelity Bayesian-optimization method designed to estimate how optimal LLM-training hyperparameters change with model and data scale. The authors tested it on synthetic benchmarks, surrogates fitted to real LLM training data and LLM pre-training runs, reporting accurate estimates with less than one-tenth the compute budget of grid search and other baselines. The results have not been peer reviewed.