Researchers propose a new approach to training large language models called Spaced Repetition Training (SRT). This method schedules the review of historical examples based on how quickly they are forgotten, using an algorithm inspired by cognitive science. SRT aims to improve the stability-plasticity trade-off in continual pre-training, allowing models to retain old knowledge while acquiring new information without erasing existing knowledge. The authors demonstrate the effectiveness of SRT on various datasets and model scales, showing improved performance compared to traditional methods. --- Why it matters: This matters because large language models are often trained on vast amounts of data, but their ability to retain old knowledge while learning new information is limited. SRT addresses this issue by providing a more adaptive approach to continual pre-training, which can improve the overall performance and stability of these models. Source: https://arxiv.org/abs/2608.17530