AI
A CMU-led team folds pretraining, fine-tuning and reinforcement learning into one algorithm
Yuda Song and co-authors, working with Johns Hopkins and Aurora, derive reinforcement-learning objectives as maximum-likelihood fits to model-generated target distributions, putting pretraining, midtraining, supervised fine-tuning and RL under a single unified training objective. Their trainer mixes static data and RL generations in one run and improves the accuracy-compute tradeoff and task coverage in the experiments reported in arXiv 2610.40564, with staged training left as a special case of choosing data, proportions and timing.