RLHF
RLHF (Reinforcement Learning from Human Feedback) is a training technique that uses human feedback to improve LLM outputs. Humans rank model responses, guiding the model toward preferred behaviors.
Definition
RLHF trains a reward model on human judgments, then uses that reward model to fine-tune the base LLM. Instead of optimizing for accuracy on a dataset, the model optimizes for generating responses that humans judge as helpful, harmless, and honest. This aligns model behavior with human preferences rather than statistical likelihood.
RLHF has made modern LLMs more useful and safer by reducing harmful outputs, improving factuality, and teaching models to acknowledge uncertainty. However, it also introduces human bias and values into models. If human raters preferred certain sources or writing styles, the model learns to favor them.
Why it matters for AI visibility
RLHF introduces human judgment into how AI search engines prioritize sources and format answers. If human raters preferred clear, well-structured explanations with specific examples, the model learns to cite sources that provide these. Understanding RLHF helps explain why content clarity and structure matter for AI visibility.
Related terms
Fine-Tuning
Fine-Tuning is the process of training a pre-trained LLM on a smaller, task-specific dataset to adapt it for particular applications or behaviors. It adjusts the model's weights after initial training.
AILarge Language Model (LLM)
Large Language Models are neural networks trained on massive text datasets to predict and generate human language. They form the foundation of modern AI search and answer engines.
AIPrompt Engineering
Prompt Engineering is the practice of crafting input text to reliably produce desired outputs from LLMs. It includes techniques for clarifying instructions, providing examples, and structuring queries.
AIAI Alignment
AI Alignment is the research field focused on ensuring AI systems behave in ways aligned with human values and intentions. It addresses how to make advanced AI systems safe and controllable.