rlhf
fact
neutral
Supervised fine-tuning alone does not guarantee task-appropriate behavior in LLMs, as the same model that achieves 88.65% accuracy on data race classification produces verbose, imprecise answers with 65.9% of responses exceeding 40 characters
Machine Learning02 Aug 2026