RLHF
Abbreviation · AI & Machine Learning · Origin: 2017
Definitions
Reinforcement Learning from Human Feedback — a training technique where human evaluators rank model outputs by quality, and the model is trained to produce outputs that align with human preferences. The process that turned GPT from 'autocomplete on steroids' into something resembling a helpful assistant.
In plain English: Training an AI by having humans rate its responses as good or bad, then teaching it to give more of the good kind — essentially teaching it manners.
Etymology
- 2017
- OpenAI and DeepMind researchers publish foundational papers on Reinforcement Learning from Human Feedback for training language models
- 2022
- InstructGPT and ChatGPT use RLHF to align model outputs with human preferences, making it the key technique behind the AI chatbot revolution
- 2023
- Alternatives like DPO (Direct Preference Optimization) emerge, but RLHF remains the industry shorthand for alignment training
Related Terms
- Fine-Tuning
- Overfitting
- Reinforcement Learning
- Alignment
- Training Data
- Classification
- Safety
- Activation Function
- AI Alignment Problem
- Direct Preference Optimization
- DPO
- Elastic Weight Consolidation
- Group Normalization
- Hidden Markov Model
- Layer Normalization
- Leaky ReLU
- Neural Scaling Laws
- Safety Alignment
- Teacher-Student Model
- Compute Optimal Training
- Neural Network Depth