Training Foundation Models with a Single Algorithm

Researchers demonstrate that foundation models do not require separate stages such as pretraining, midtraining, supervised fine-tuning, and reinforcement learning. By deriving RL objectives as maximum-likelihood fitting to model-generated target distributions, they unify these lifecycle stages into a single likelihood-based training algorithm. Across experiments, this unified training approach improves the accuracy-compute tradeoff, enhances task coverage, and provides a stronger starting point for continual learning. Open article →