Models can get caught by what they start to say early. So if they model goes down a path that seems like a likely answer early on, and that ends up being a false lead or dead end, they will end up making up something plausible sounding to try and finish that line of thought even if it's wrong. This is why chain of thought and other "pre-answer" techniques improve results.

Because of the way transformers work, they have very good hindsight, so they can realize that they've just said things that are incorrect much more often than they can avoid saying incorrect things.

Sorry for the late reply, but if you read this, there is research that shows that prompting a LLM to take variety of perspectives on a problem (IIRC it was demonstrated with code) then finding the most common ground answer improved benchmark scores significantly. So, for example if you ask it to provide a brief review and likelihood of the answer, and repeat that process from several different perspectives, you can get some very solid data.