rlhf
fact
bearish
RLHF preference tuning drives LLMs to overuse emphatic rhetorical figures like epanorthosis, with the training distribution rich in promotional prose being the main driver rather than the left-to-right generation process
Computation and Language29 Jul 2026