Techniques

RLHF (Reinforcement Learning from Human Feedback)

Training technique that aligns LLMs to human preferences using reinforcement learning on preference data.

RLHF trains a reward model on human preferences (response A preferred over response B), then uses reinforcement learning to update the LLM to maximise reward. It's how ChatGPT, Claude and other frontier assistants became helpful, harmless and honest.

In 2024–26, simpler alternatives like Direct Preference Optimization (DPO) and its variants (IPO, KTO) have largely replaced classic RLHF for enterprise fine-tuning. DPO achieves similar alignment with much less complexity — no separate reward model, no RL loop.

For most enterprise teams, preference tuning via DPO is the practical path. Classic RLHF remains largely a frontier-lab technique.

Key points

  • Aligns models to human preferences via reward + RL
  • DPO and variants have replaced classic RLHF for most enterprise use
  • Requires preference data (harder to collect than SFT data)
  • Best for subjective quality — tone, helpfulness, safety

Common use cases

Aligning consumer-facing assistants to brand tone
Safety and refusal calibration
Lifting subjective quality beyond SFT ceiling

Frequently asked

Related terms

Fine-Tuning

Continuing training of a pre-trained model on domain-specific data to specialize its behavior.

Large Language Model (LLM)

A neural network trained on massive text corpora to understand and generate human-like language.

Applying RLHF (Reinforcement Learning from Human Feedback) in your business?

We help enterprises design, ship, and govern AI systems end-to-end.

Start AI Readiness Check Talk to an expert
Back to glossary