The research team from National University of Defense Technology systematically described the memory efficiency issues in the training process of large language models (LLMs) with Transformer architecture for scientific fields. As the scale of models continues to expand, memory requirements have become a key bottleneck restricting their applications in scientific fields such as biology, medicine, chemistry, and meteorology. The article systematically reviewed memory-efficient pre-training technologies at the algorithm level, system level, and software-hardware collaborative optimization, and combined with the AlphaFold 2 case, demonstrated how customized memory optimization methods can reduce storage overhead while maintaining prediction accuracy.

Finding a few tumor cells hidden among billions of blood cells is one of the biggest challenges in liquid biopsy. Circulating tumor cells (CTCs), which break away from primary tumors and enter the bloodstream, can provide important clues about cancer progression and treatment response. However, their extremely low abundance—often only one to ten cells among a billion blood cells—makes them difficult to isolate and accurately identify.

In a new study published in Biomedical Analysis, researchers from Taiyuan University of Technology and Shanxi Bethune Hospital developed a new strategy that combines a microfluidic chip with artificial intelligence to address this challenge. The platform first separates tumor cells from surrounding blood cells according to their physical differences, then uses deep learning to identify the enriched cells from ordinary microscope images. This approach provides a new route toward simpler, less label-dependent CTC analysis.