True intelligence knows when it's presented with a proposition, whether or not the proposition contradicts one or more of its existing beliefs.

LLMs can be trained on arbitrary token strings, which semantically contradict each other, because semantics is not involved.

The way to attack the "garbage out" problem with an LLM (like anything else) is to address the "garbage in" problem.

You will get better quality out of an LLM that is trained on nothing but mountains of unvarnished truth: logically sound propositions and true facts about the world.

But even then, interpolating among such is not reliable! We can't add together two different truths and divide by two to obtain a midpoint truth: that makes no sense.

That is still a sort of "garbage in" problem. But that one is not in the data; it is in the design of the processing concept. The body of algorithms that power LLMs have a fundamental garbage aspect to them that cannot be removed; only a replacement of the LLM with something different can address it.