EasyDeepLearn

RLAIF — RL from AI Feedback.

hard

Answer

  • Replace human labelers with LLM judge for preference labels.
  • Bai et al. 2022 (Constitutional AI): iteratively use a strong LLM to critique + revise responses per a written constitution.
  • Advantages: scale (billions of preferences), consistency, cost.
  • Trade-offs: risk of amplifying LLM biases, less alignment with real user preferences.
  • Standard in modern alignment (Anthropic's Constitutional AI, Meta's Llama, Google's Gemini all use RLAIF + human).
Check yourself — multiple choice
  • Random
  • Replace human preference labels with LLM judge (per constitution); scale + consistency + cost; risk of amplifying LLM bias; used in Constitutional AI / Llama / Gemini + human hybrid
  • Same as RLHF
  • Not real

RLAIF: LLM judge for preferences; Constitutional AI; scales beyond humans.

#llm#alignment

Practise Reinforcement Learning

214 interview questions in this topic.

Related questions