Evidence-first authority page
What Is RLHF (Reinforcement Learning from Human Feedback)?
A definition of RLHF, why it's used to align model behavior beyond raw next-token prediction, and its practical effect on output quality.
Direct answer
RLHF (Reinforcement Learning from Human Feedback) is a training technique where a model is further trained using human preference rankings of its outputs, rewarding responses humans rate as more helpful, honest, or safe — this is a major reason modern chat models follow instructions and refuse harmful requests, behaviors not directly present in raw next-token-prediction training.
Last reviewed 2026-08-29
This page remains noindex while its page-level source receipts are completed and verified. It is available for review, not presented as publication-ready authority.
Why base training alone isn't enough
A model trained purely to predict the next token from internet-scale text learns to continue text plausibly, but has no inherent notion of being helpful, following instructions precisely, or declining an inappropriate request — those behaviors come from additional training stages, of which RLHF (or related techniques) is typically one.
How the process generally works
- Human reviewers rank multiple model outputs for the same prompt by preference.
- A reward model is trained to predict these human preference rankings.
- The base model is further trained (via reinforcement learning) to produce outputs the reward model scores highly.
What this means practically for users
A model's tendency to be conversational, follow formatting instructions, and refuse certain requests is shaped substantially by this post-training process, not solely by its base training data — this is part of why two models with similar base training can behave quite differently in practice, and why a vendor's specific alignment choices are a real, distinct product characteristic worth considering separately from raw capability.