I read a research paper which says LLMs don’t actually improve by training on the outputs of other LLMs. They only gain the ability to answer the specific ques

I read a research paper which says LLMs don’t actually improve by training on the outputs of other LLMs. They only gain the ability to answer the specific questions included in the LLM output training dataset. Their reasoning capabilities etc. don’t improve.