Researchers introduce Trajectory Soup, a method to distribute mid-training budgets over multiple independent branches of large language models, improving aggregate downstream performance and extending the compute-scaling frontier of mid-training.