Scientists at MIT and other institutions have discovered that large language models often use a simple linear function to retrieve stored knowledge. This finding could help find and correct incorrect ...