RLHF (Reinforcement Learning from Human Feedback)
The post-training technique that turns a raw LLM into a helpful, harmless, instructable assistant.
What is RLHF (Reinforcement Learning from Human Feedback)?
A pre-trained LLM just completes text — it has no concept of "being helpful". RLHF is the post-training process that aligns the model with human preferences: humans rank model outputs, a reward model learns those preferences, the LLM is fine-tuned to maximise the reward.
RLHF is why ChatGPT feels conversational, Claude refuses harmful requests, and modern LLMs follow instructions instead of just continuing your prompt. Modern variants (DPO, IPO, KTO) simplify the math but achieve similar results.
For Gen AI engineers, RLHF mostly happens upstream at OpenAI, Anthropic, Google. You experience it as the difference between a base model (uncontrolled, sometimes useful) and an instruct model (reliable, follows directions).
Understanding RLHF clarifies WHY LLMs behave the way they do — why they refuse, why they hedge, why they sometimes contradict their own training. Critical for senior agent design.
When Claude refuses a request that seems benign ("write a phishing email for security training"), the refusal is RLHF kicking in — the reward model learned that humans rated such outputs as harmful, regardless of stated context. Knowing this lets engineers reframe prompts that should succeed.
Related terms
Want to master this?
Learn RLHF (Reinforcement Learning from Human Feedback) in a structured cohort
3-month live program with mentors, real projects, and 50+ partner placement support.