Alibaba and Renmin University open-source LOGOS, a science model that beats Microsoft's NatureLM with fewer parameters

Alibaba and Renmin University of China have open-sourced LOGOS, a new model for scientific discovery that uses a unified 'scientific grammar' to understand and generate data across different fields like proteins, molecules, and materials.

LOGOS handles six major scientific tasks using a simple sequence-based approach, matching or beating specialized methods. Its 1-billion-parameter version outperforms Microsoft's NatureLM (which has 8x7B parameters) on several tasks, showing incredible efficiency. It also uses a clever 'text description' method to understand complex 3D interactions without needing actual 3D coordinates.

The model was trained on a huge dataset of 44.87 billion tokens covering 7 types of scientific data, including proteins, antibodies, and small molecules. Because its training and real-world tasks use the same format, LOGOS works well without lots of extra fine-tuning. All the model weights, code, and research paper are now available online.