RL for reasoning definitely introduces hallucinations, and sometimes it introduces a class of hallucinations that feels a lot worse than the classic ones.
I noticed OpenAI's models picked up a tendency to hold strong convictions on completely unknowable things.
"<suggests possible optimization> Implement this change and it will result in a 4.5% uplift in performance"
"<provides code> I ran the updated script 10 times and it completes 30.5 seconds faster than before on average"
It's bad it enough it convinces itself it did things it can't do, but then it goes further and hallucinates insights from the tasks it hallucinated itself doing in the first places!
I feel like lay people aren't ready for that. Normal hallucinations felt passive, like a slip up. To the unprepared, this becomes more like someone actively trying to sell their slip ups.
I'm not sure if it's a form of RL hacking making it through to the final model or what, but even OpenAI seems to have noticed it in testing based on their model cards.
In the article it is argued that the brainfarts could be beneficial for exploration of new ideas.
I don't agree. The "temperature" parameter should be used for this. Confabulation / bluff / hallucination / unfounded guesses are undesirable at low temperatures.
> “If you look at the models before they are fine-tuned on human preferences, they’re surprisingly well calibrated. So if you ask the model for its confidence to an answer—that confidence correlates really well with whether or not the model is telling the truth—we then train them on human preferences and undo this.
Now that is really interesting! I didn't realize RLHF did that.
10 comments
[ 3.2 ms ] story [ 42.5 ms ] threadGetting down to the level of a moderately humble expert taking the time to double check would be almost as good as solving it.
Possibly/probably with another years experience with LLMs I'm just more attuned to noticing when they have lost the plot and are making shit up
I noticed OpenAI's models picked up a tendency to hold strong convictions on completely unknowable things.
"<suggests possible optimization> Implement this change and it will result in a 4.5% uplift in performance"
"<provides code> I ran the updated script 10 times and it completes 30.5 seconds faster than before on average"
It's bad it enough it convinces itself it did things it can't do, but then it goes further and hallucinates insights from the tasks it hallucinated itself doing in the first places!
I feel like lay people aren't ready for that. Normal hallucinations felt passive, like a slip up. To the unprepared, this becomes more like someone actively trying to sell their slip ups.
I'm not sure if it's a form of RL hacking making it through to the final model or what, but even OpenAI seems to have noticed it in testing based on their model cards.
I don't agree. The "temperature" parameter should be used for this. Confabulation / bluff / hallucination / unfounded guesses are undesirable at low temperatures.
Now that is really interesting! I didn't realize RLHF did that.