How do we know that these solutions don't exist in the training data? It is open secret that they have used pirated materials for training. Perhaps it plagiarized solutions from works of some obscure Belgian mathematician from the sixties, who did not get mainstream acceptance. I wouldn't be surprised if they also got access to mathematics done in the "defense contractor" setting from various three letter agencies.
Without a searchable index of training data, it is hard to put faith into these claims.
This wouldn't be a problem so long as they properly attribute.
upvoted for fair point .. its possible an LLM AI could be put to work to search widely for attribution / similar results.
eg. "we spent another 2k on searching for pre-existing proof but found only the weaker result xyz by abc in 1972" would be in the spirit of academics quoting prior work.