RLHF is the post-training technique that aligned raw language models into helpful assistants: humans rank candidate outputs, a reward model learns those prefere
Reinforcement Learning from Human Feedback is the post-training technique that aligned raw language models into helpful assistants: humans rank candidate outputs, a reward model learns those preferences, and the model is optimized to produce preferred responses.
Known failure modes include sycophancy and reward hacking, which is why alignment work continues beyond it. RLHF remains foundational but is now extended by AI-feedback variants (RLAIF) and newer optimization methods.