NYU CS-GY 9223 L: Language Models, Reinforcement Learning, Reasoning (Fall 2026)

Professor Pavel Izmailov

Thursday 2:00-4:30pm, Jacobs Hall, 6 Metrotech, Room 204

Course Description

This is a graduate seminar-style course on modern large language models (LLMs), reinforcement learning (RL), and reasoning. The goal is to bring students to the frontier of the field: we will read and discuss contemporary research papers and technical reports, and work through the engineering and scientific decisions behind today's state-of-the-art systems.

Evaluation is primarily project-based. There are no written exams.

Professor Contact

Email: pi390@nyu.edu

Course Assistants

Alex N. Wang

Links

Lectures

Date	Topic	Main Reading	Additional Reading
09/10/2026 Slides Recording	• LLM basics	• Attention Is All You Need	• GPT-2: Language Models are Unsupervised Multitask Learners • GPT-3: Language Models are Few-Shot Learners
09/17/2026	• RL foundations: MDPs • Policy gradients • PPO	• Sutton & Barto — Reinforcement Learning: An Introduction	• PPO — Proximal Policy Optimization Algorithms • DQN — Playing Atari with Deep Reinforcement Learning • GAE — Generalized Advantage Estimation
09/24/2026	• Scaling laws	• Kaplan et al. — Scaling Laws for Neural Language Models • Chinchilla — Training Compute-Optimal Large Language Models	• Scaling Data-Constrained Language Models
10/01/2026	• AlphaGo / AlphaZero	• AlphaZero — Mastering Chess and Shogi by Self-Play	• MuZero — Mastering Atari, Go, Chess and Shogi with a Learned Model
10/08/2026	• Attention variants and MoE	• DeepSeek-V3 (architecture sections only)	• GQA — Grouped-Query Attention • DeepSeekMoE
10/15/2026	• Positional encoding and long context	• RoPE — RoFormer • YaRN — Efficient Context Window Extension	• RULER — What's the Real Context Size of Your Long-Context Language Models?
10/22/2026	• Optimizers and training dynamics	• Muon is Scalable for LLM Training (Moonlight)	• muP — Tensor Programs V • Keller Jordan — Muon (blog post)
10/29/2026	• RLHF and preference optimization	• InstructGPT — Training Language Models to Follow Instructions • DPO — Direct Preference Optimization	• Constitutional AI • Tülu 3
11/05/2026	• DeepSeek-R1 • GRPO • RLVR	• DeepSeek-R1	• DeepSeekMath — GRPO derivation • Kimi k1.5 • Chain-of-Thought Prompting (background)
11/12/2026	• Reward hacking and over-optimization	• Gao, Schulman & Hilton — Scaling Laws for Reward Model Overoptimization	• Baker et al. — Monitoring Reasoning Models and the Risks of Promoting Obfuscation • Concrete Problems in AI Safety
11/19/2026	• Weak-to-strong generalization	• Burns et al. — Weak-to-Strong Generalization	• Turpin et al. — Language Models Don't Always Say What They Think • AI Safety via Debate
12/03/2026	• Kimi K3, full technical report	• Kimi K3 technical report	• Kimi Linear (KDA) • Kimi K2
12/10/2026	• World models	• V-JEPA 2	• DreamerV3 • GameNGen — Diffusion Models Are Real-Time Game Engines • Genie 3 (DeepMind blog post)