There are two reasons:

1. LLMs aren't run deterministically - they use a bit of noise to improve the output. So even with the exact same input you're going to get a whole distribution of outputs.

2. Feedback from the environment. LLMs (and humans!) work best when they can test ideas. That's why coding agents work so well. If you ask an agent 1000 times to just think of a compression algorithm, it'll probably spit out very samey answers just like a human would. If you give it a test corpus and let it try out different things it will absolutely be able to come up with novel ideas.