Why Do Task Vectors Exist in Pretrained LLMs? This AI Research from MIT and Improbable AI Uncovers How Transformers Form Internal Abstractions and the Mechanisms Behind in-Context Learning (ICL)  MarkTechPost