A newly released machine learning method seeks to eliminate the traditional boundaries between the different phases of building artificial intelligence foundation models. The system, called Unified Training, merges pretraining, midtraining, supervised fine-tuning, and reinforcement learning (RL) into a single maximum-likelihood objective, which allows developers to train static and RL-generated data within the same framework.
Building on the MaxRL method, the approach aims to simplify the infrastructure required for model development and lower compute costs. The researchers state that this unified process reduces the likelihood of models forgetting previous information and enhances their ability to solve complex reasoning tasks compared to traditional staged pipelines.
