Hacker News Reader

> Problems near identical to the test set can be found online.

I don’t see how that is relevant to a model trained with RL. If it had already memorized the answers then why was any training necessary?