MIT researchers discovered the underlying cause of position bias, a phenomenon that causes large language models to overemphasize the beginning or end of a document or conversation, while neglecting the middle. They built a theoretical framework that can be used to diagnose and correct position bias in future model designs, leading to more accurate, reliable AI agents.

A low-cost, scalable fabrication technology developed at MIT can integrate fast, efficient gallium nitride transistors onto a standard silicon chip, which could boost the performance of electronic chips used in high-bandwidth applications like video calling and real-time deep learning.