RLAIF
Reinforcement learning from AI feedback instead of human feedback.
Overview
RLAIF replaces the human labellers in RLHF with a model that judges outputs against given criteria. It solves the bottleneck that human preference data is slow and expensive to collect at the scale alignment training needs. It exists because the judge can be given explicit principles and applied consistently, and the open question it leaves is whether a model's judgement is a good enough substitute for a person's.
Technology Graph
How RLAIF connects to the rest of the stack.
Uses and depends on
- LLM-as-a-Judge technique Evaluation RLAIF uses LLM-as-a-Judge High ✓ source
- RLHF technique Techniques RLAIF based on RLHF High ✓ source
Applied to
- Alignment technique Safety & Alignment RLAIF used for Alignment High ✓ source