LLMs don't. There are tools that can fetch new information and then gets fed into the model as more tokens, but that's just a special case of what kescusay is saying about injecting text.