> There's no "True" to an LLM, just how probable tokens are given previous context.

It may be enough: tool assisted LLMs already know when to use tools such as calculators or question answering systems when hallucinating an answer is likely to impact next token prediction error.

So next-token prediction error incentivize them to seek for true answers.

That doesn't guaranty anything of course, but if we were only interested in provably correct answers we would be working on theorem provers, not on LLMs