LLMs show human-like representational change during associative learning, driven by pair similarity and modulated by prior knowledge.