MLFine-TuningAgentic

CS 185/285: Deep Reinforcement Learning (UC Berkeley, Spring 2026)

by UC Berkeley (Sergey Levine, RAIL)

AdvancedCourseFree~14 weeks, 25+ lectures plus 5 programming homeworks; self-paced from the public slides and the Fall 2023 lecture recordings

The canonical deep RL course, now with an LLM RL lecture, homework and final-project track.

Start LearningAdded Sep 16, 2026 · Updated Sep 16, 2026

Overview

The Spring 2026 offering runs roughly fourteen weeks and twenty-five-plus lectures, in a fixed progression: course intro and imitation learning (lectures 1-3), RL basics (4), policy gradients and actor-critic (5-6), value-based RL (7-8), advanced policy gradients (9-10), variational inference (11-12), control as inference and LLM RL (13-14), model-based RL (15-16), offline RL (17-18), exploration and RL theory (19-20), advanced exploration and multi-task RL (23-24), then open problems and guest lectures. Five programming homeworks carry it: HW1 imitation learning, HW2 policy gradients, HW3 Q-learning and actor-critic, HW4 LLM RL, HW5 offline RL. The public grading breakdown is homeworks 50 percent (10 percent each), final project 20 percent, exam 20 percent, mini-quiz 10 percent — a full workload, not a survey. The syllabus states plainly that CS189 or equivalent is a prerequisite and that the course assumes familiarity with reinforcement learning, numerical optimization and machine learning, recommending the CS188 EdX intro RL material and Sutton & Barto chapters as review. Lecture slides are downloadable as PDFs and the assignments live in the MIT-licensed homework_spring2026 repository (updated May 2026, ~141 stars), which also contains two final-project scaffolds: final_project_llm_rl and final_project_offline_online. Lecture recordings from the Fall 2023 offering are on YouTube; Spring 2026 recordings are listed as TBD. Sergey Levine's RAIL lab produced much of the offline and model-based RL literature the course teaches.

At a Glance

Topic
ML
Level
Advanced
Format
Course
Cost
Free
Duration
~14 weeks, 25+ lectures plus 5 programming homeworks; self-paced from the public slides and the Fall 2023 lecture recordings
Provider
UC Berkeley (Sergey Levine, RAIL)
Hands-on
Yes — code/exercises
Certificate
None

What You’ll Learn

  • Implement behavior cloning and imitation learning agents from scratch in Homework 1
  • Derive and implement policy gradient methods, then extend them with actor-critic baselines
  • Build Q-learning and actor-critic agents and reason about value-based deep RL failure modes
  • Apply reinforcement learning directly to language models in the dedicated LLM RL assignment
  • Train and evaluate offline RL agents that learn from fixed datasets without new interaction
  • Work through model-based RL, exploration strategies and the RL theory lectures
  • Connect control-as-inference and variational inference to the RL objectives used in post-training

Highlights

  • The Spring 2026 edition adds an LLM RL lecture, an LLM RL homework and an LLM RL final-project track — still rare in a classical RL course
  • Everything is public: lecture slide PDFs on the course site, assignments in the MIT-licensed homework_spring2026 repo
  • Taught by Sergey Levine of Berkeley's RAIL lab, a central figure in the offline and model-based RL work the syllabus covers
  • The grading breakdown is published (50% homework, 20% project, 20% exam, 10% quizzes), so a self-study learner knows the real effort split
  • Full lecture recordings from the Fall 2023 offering are on YouTube for anyone following along outside Berkeley

Who It’s For

Best For

  • Engineers moving from supervised fine-tuning into RL post-training and wanting the algorithms, not the wrapper
  • ML researchers who need full derivations of policy gradient, value-based and offline RL methods
  • Practitioners who already run TRL, verl or NeMo-RL and want to understand what those libraries implement
  • Robotics and control engineers applying deep RL to real decision-making systems

Prerequisites

  • Berkeley CS189 or an equivalent graduate-level machine learning course (stated prerequisite)
  • Assumed familiarity with reinforcement learning basics, numerical optimization and machine learning
  • Strong Python and a deep learning framework for the five programming assignments
  • Optional review: the CS188 EdX intro RL material and the recommended Sutton & Barto chapters

FAQ

What is CS 185/285: Deep Reinforcement Learning (UC Berkeley, Spring 2026)?

UC Berkeley's graduate deep reinforcement learning course, taught by Sergey Levine and currently offered as CS 185/285 in Spring 2026. It is the reference RL curriculum for engineers who suddenly need to post-train language models: the Spring 2026 edition carries a dedicated LLM RL lecture, an LLM RL programming assignment and an LLM RL final-project option alongside the classic imitation learning, policy gradient, value-based, model-based and offline RL material. Slides and the homework repository are public.

Is CS 185/285: Deep Reinforcement Learning (UC Berkeley, Spring 2026) free?

CS 185/285: Deep Reinforcement Learning (UC Berkeley, Spring 2026) is free to access.

What level is CS 185/285: Deep Reinforcement Learning (UC Berkeley, Spring 2026) for?

CS 185/285: Deep Reinforcement Learning (UC Berkeley, Spring 2026) is aimed at a advanced audience. Recommended background: Berkeley CS189 or an equivalent graduate-level machine learning course (stated prerequisite), Assumed familiarity with reinforcement learning basics, numerical optimization and machine learning, Strong Python and a deep learning framework for the five programming assignments, Optional review: the CS188 EdX intro RL material and the recommended Sutton & Barto chapters.

How long does CS 185/285: Deep Reinforcement Learning (UC Berkeley, Spring 2026) take?

Expect roughly ~14 weeks, 25+ lectures plus 5 programming homeworks; self-paced from the public slides and the Fall 2023 lecture recordings. Most learners work through it at their own pace.

What will I learn from CS 185/285: Deep Reinforcement Learning (UC Berkeley, Spring 2026)?

You'll learn: Implement behavior cloning and imitation learning agents from scratch in Homework 1; Derive and implement policy gradient methods, then extend them with actor-critic baselines; Build Q-learning and actor-critic agents and reason about value-based deep RL failure modes; Apply reinforcement learning directly to language models in the dedicated LLM RL assignment; Train and evaluate offline RL agents that learn from fixed datasets without new interaction; Work through model-based RL, exploration strategies and the RL theory lectures; Connect control-as-inference and variational inference to the RL objectives used in post-training.

Topics

deep reinforcement learningllm rlpolicy gradientsoffline rlpost-training

Sources

This page was written from 4 sources, 2 on domains other than rail.eecs.berkeley.edu.

  1. 1.rail.eecs.berkeley.edudeeprlcoursevendor
  2. 2.rail.eecs.berkeley.edusyllabusvendor
  3. 3.github.comhomework spring2026
  4. 4.github.comberkeleydeeprlcourse