Predictable Scaling Laws of Optimal Hyperparameters for LLM Continued Pre-training

In three linesarXiv paper demonstrating that optimal hyperparameters for LLM continued pre-training follow predictable scaling laws. Two-stage framework: empirical law discovery via proxy models, then state-aware prediction using validation loss and equivalent pre-training compute. Reduces hyperparameter search overhead by 90% while maintaining performance.

Summary generated by Claude — human-verified