LLMs are famously bad at determining "if it's not sure it is correct." They are always confident, because a confident tone ranks better in RL.
This makes it sound like RL rewards a confident tone -- in general, I don't think this is true (most RL is RLVR, which typically uses binary verification of correctness).
I say this because the real reason "they are always confident" is in some sense even more contrived. Training text where the speaker sounded more confident is more likely to contain a correct answer.
Generally it does. Especially in groups. Hell look at the state of politics right now: it’s basically about being the loudest, least compromising, most confident voice in the room. It’s not just because people will assume you’re correct, it’s because if you are confidently saying something that someone wants to be right, then they’re often just going to follow it. We are all guilty of this.
If I’m turning to an LLM to diagnose something medical, I am probably frustrated or uncomfortable. Maybe I’m just scared. So this magic device just instantly spits out (allegedly) exactly what is wrong and exactly what I need to do with no hesitation. I am very liable to just take it at face value because I want an answer and it gave me one, as we have seen over and over again since ChatGPT was unleashed on the world.
We don’t really need to speculate, this is already a problem.
I was unintentionally being pedantic, because this isn't really done with RL anymore - it doesn't need to be. RL is now typically only used to train reasoning for tasks with a well-defined correct answer (that's what I meant by binary reward) - this is called RLVR (RL with verifiable rewards).
Preference optimization (training the model on user "this response is better than that response" type data) is more often done with something in the same family as DPO (direct preference optimization), which is decidedly not RL.
Your philosophical concerns are correct of course. And there's the added caveat that the models that most people are using are closed, so we don't actually know their training recipes for sure.
A binary response vs rating is not related whether it learns confident or hedged tone. Either will produce a confident tone because humans respond more positively to a confident tone, hence the conman's language. Binary or not humans reward the tone and very much bias the model.
But there's an even more contrived reason the training set contributes. The vast majority of human writing is confident. When the prior is greatly biased, a random number generator biased to that prior does better. The difference with humans and machines is humans are less likely to respond if they are less confident because they understand not knowing, which is why the training set is biased. It is one of the many fundamental flaw of LLM training and confusion of LLMs with intelligence. And that will not be fixed within the LLM architecture.
Yes, I am aware. I was referring to the fact that at this point, user preference optimization is not done with RL, but with other techniques. I was unintentionally being pedantic.
> But there's an even more contrived reason the training set contributes. The vast majority of human writing is confident.
Yes, I think this has more to do with it than user preference optimization, honestly. "Valuable" text (i.e. text that produces a "good" model) for pretraining has the characteristic of being confident. Even models which are not optimized for chat (i.e. definitely no user preference data used to train them) exhibit this characteristic for medical questions (I know, because I've literally tested them for this purpose).
Preference optimization (or even RLVR) might play some small role as well, but it's kind of a "turtles all the way down" type problem.