The findings, which apply to both large language models (LLMs) and vision language models (VLMs), challenge one of the main beliefs of the LLM community — that models require hand-labeled ...