The challenge is more about providing actual, objective data to learn this.
The LLM will not be able to decide what is "good UX" and "bad UX", it will have to ingest this information in vast detail from judgement provided by humans, which is also varying alot.
Case in point: Coming from that field, I personally don't have a good user experience using iOS, a majority of end-users will strongly disagree.
An LLM would need to be trained to handle both cases to judge the same human interface, so it would need alot of data to understand both views...
(also: UX is not UI, but I guess this goes without saying...)