What is the key idea behind Reinforcement Learning from Human Feedback (RLHF) in LLM alignment?
-
A
Fine-tuning the model on synthetic data generated by a rule-based reward function
-
B
Training a reward model on human preferences, then using RL to optimize the LLM against that reward
-
C
Replacing next-token prediction with binary yes/no feedback signals
-
D
Using RLHF only during the pre-training phase to reduce harmful outputs