You used both phrases in this thread, but those are two very different things. It's a stretch to say this research supports the latter.
Yes, LLMs are still next-token generators. That is a descriptive statement about how they operate. They just have embedded knowledge that allows them to generate sometimes meaningful text.