Reinforcement Learning from Human Feedback

A short introduction to RLHF and post-training focused on language models.

Nathan Lambert

Course

A full course accompanying the book with added resources and other lectures I've given.

The slide decks are usually built with Claude Opus, Colloquium, and substantial human revisions.

Welcome to the Course

Introduction and overview of what you'll learn

Prerequisites

This course is roughly aimed at early AI PhD or master's students, but it is designed to be accessible to anyone willing to put in the work.

You do not need prior reinforcement learning or language modeling background to start. A motivated learner who studies hard — leaning on today's top LLMs as a tutor to unpack unfamiliar math, code, and jargon — can follow the entire course. I encourage you to go down rabbit holes, skip or reorder videos, and chase what excites you.

If you prefer the traditional coursework path, the usual background is the basics of language modeling / NLP plus basic machine learning (e.g. an intro to AI course and an intro to ML course).

Lecture 0: The ML Foundations of LLM Post-Training

A refresher on the ML prerequisites of post-training — language modeling, KL, cross-entropy, & other math — to acclimate to the series

Additional learning material

Primary Material

Lecture 1: Overview

Chapters 1-3 · Foundations of RLHF and post-training

Lecture 2: IFT, Reward Models, & Rejection Sampling

Chapters 4, 5, 9 · Start of the core optimization methods section

Lecture 3: RL Motivation & Math

Chapter 6, Part 1 · Policy gradients math, intuitions, and theory

Lecture 4: RL Implementation & Practice

Chapter 6, Part 2 · Code, loss aggregation, async training, and practical engineering

Lecture 5: The Rise of Reasoning Models

Chapter 7 · RLVR, inference-time scaling, and the 2025 reasoning model wave

Lecture 6: Direct Preference Optimization

Chapter 8 · Deriving DPO step by step, plus variants and practice

Conversation 1: Frontier post-training recipes in 2026 (w/ Finbarr Timbers)

Lecture 7: Synthetic Data and Modern Post-training Methods

Chapter 12 · On-policy distillation, AI feedback, Constitutional AI, and rubrics

Lecture 8: On "Preferences" and Preference Data

Chapters 10 & 11 · Where the idea of quantifying preference came from and opaque data practices

Lecture 9: Over-Optimization and RLHF's Bad Reputation

Chapter 14 & Appendix B · Goodhart's law, reward hacking, sycophancy, and style

Lecture 10: Regularization in RL, Why RL Generalizes, and Why SFT Forgets

Chapter 15 · The KL penalty, implicit regularization, and why RL forgets less

Conversation 2: From academic research to frontier practice (w/ Scott Geng)

We discuss what it takes to land a well-grounded academic result into a near-frontier model. Scott and I worked together on DPO for Olmo 3, and how building models is much more than the idea, it's making it fit in a more complex puzzle.

Lecture 11: Tool Use, Function Calling and The Road to Agents

Chapter 13 · Why models need tools, MCP and harnesses, and tool-use RL

Lecture 12: The Evolution of Frontier Model Evaluation

Chapter 16 · The eras of evaluation, agentic evals, and trusting the numbers

Lecture 13: An Introduction to Character Training

Chapter 17 · Constitutions, model specs, model personalities, and open questions

Extra Resources

Outside material that I've personally used for going deeper on reinforcement learning and language models. The books in particular are wonderful complements.

Books & Courses

More talks from Nathan

Various presentations over my last few years working in post-training, grouped by corresponding chapter.

Other Guest Lectures and Talks

2026

An Introduction to Reinforcement Learning from Human Feedback and Post-training

SALA 2026 · Quito, Ecuador · March 2026

Citation

If you found this useful for your research, please cite it!

For the web and arXiv version:

@misc{lambert2025reinforcementlearninghumanfeedback,
  title = {Reinforcement Learning from Human Feedback},
  author = {Nathan Lambert},
  year = {2025},
  eprint = {2504.12501},
  archivePrefix = {arXiv},
  primaryClass = {cs.LG},
  url = {https://arxiv.org/abs/2504.12501}
}

For the Manning edition:

@book{lambert2026reinforcement,
  author = {Nathan Lambert},
  title = {Reinforcement Learning from Human Feedback: Alignment and post-training of {LLMs}},
  year = {2026},
  publisher = {Manning Publications},
  isbn = {9781633434301},
  url = {https://www.manning.com/books/reinforcement-learning-from-human-feedback}
}