Pretraining, fine-tuning and reinforcement learning can share one training algorithm, researchers show
A new paper recasts pretraining, midtraining, supervised fine-tuning and suitable reinforcement learning as one maximum-likelihood objective. Every source, from static text to the model's own RL attempts, becomes weighted examples for a single trainer, so the design question becomes which data to mix, and when. Staged training is one special case. In their tests, mixing reduced forgetting and reached higher accuracy and coverage at larger compute budgets.