5 ms·
Exactly, I think that by their very design, LLMs are very sensitive to how a question is framed. But I wonder how much of that comes from RLHF itself or just f
by stonecauldron 6mo ago
Exactly, I think that by their very design, LLMs are very sensitive to how a question is framed.
But I wonder how much of that comes from RLHF itself or just from the way token prediction works.
- rzmmm 6mo agoIt's likely the RLHF process since there are significant differences between models about this.