CS 329X: Human-Centered LLMs (Stanford, Fall 2026)
by Stanford University (Diyi Yang)
Stanford's research course on preference tuning, human-AI interaction, evaluation and safety for LLM systems
Overview
CS 329X: Human-Centered LLMs is a Stanford graduate course taught in Fall 2026 by Diyi Yang, with Yijia Shao as head TA and three further TAs. It combines NLP and HCI perspectives and teaches methods for designing, aligning, personalizing and evaluating LLM systems that serve diverse users. The schedule opens on September 23 with an introduction to human-centered LLMs, then a crash course on LLMs built around DSPy and a survey of prompting techniques, then learning from human preferences through the InstructGPT and DPO papers. Week 3 covers data (LIMA, FineWeb, data authenticity) with guest speakers Will Held and Rafal Wojdyla, followed by human-AI interaction. Weeks 4 and 5 turn to evaluating human-AI interaction, human-agent collaboration with Collaborative Gym, building full-stack human-AI systems, and generative and elicitive interfaces. Later weeks cover LLMs and the future of work, AI companions and the INTIMA benchmark, safety and privacy in LLM use, and a guest lecture on overreliance and sycophancy, ending with final presentations on December 2. Enrolled students complete three homeworks and a quarter-long project with a proposal, a 15-minute video, a midway showcase and a final report due December 9. Lecture slides are posted as PDFs on the course site. The week-one reading is a May 2026 arXiv paper by Caleb Ziems and 57 co-authors, Yang among them, with recommendations for human-centered design, data sourcing, training, evaluation and deployment. Prerequisites are CS224N or CS229, or equivalent background.
At a Glance
- Topic
- Models
- Level
- Advanced
- Format
- Course
- Cost
- Free
- Duration
- 10-week quarter (Sep 23 to Dec 2, 2026), two lectures a week; public slides and readings, self-paced
- Provider
- Stanford University (Diyi Yang)
- Hands-on
- No
- Certificate
- None
What You’ll Learn
- ✓How RLHF and direct preference optimization turn human preferences into model behavior
- ✓How data choices such as LIMA-style curation and FineWeb filtering shape model quality
- ✓Methods for evaluating human-AI interaction beyond static benchmark accuracy scores
- ✓How to study and design human-agent collaboration, including the Collaborative Gym setup
- ✓Design approaches for full-stack human-AI systems and generative, elicitive user interfaces
- ✓How LLM adoption changes work, drawing on automation versus augmentation research
- ✓Risks in LLM use including privacy, safety, overreliance, sycophancy and AI companionship
- ✓Programming LLM pipelines with DSPy and comparing prompting techniques from a systematic survey
Highlights
- •The full Fall 2026 reading list is public, with each lecture tied to specific papers such as InstructGPT, DPO, LIMA and FineWeb
- •Lecture slides are posted as PDFs, so you can follow along without enrolling
- •Spends several weeks on interaction evaluation, human-agent collaboration, overreliance and sycophancy, topics that sit between NLP and HCI
- •Guest lectures from researchers including Emma Pierson, Valerie Chen, Weiyan Shi and Kunal Handa
- •The week-one reading is a 58-author May 2026 arXiv paper on human-centered LLMs that Yang co-authored
Who It’s For
Best For
- ✓ML engineers working on preference tuning or post-training data
- ✓Engineers designing agent UX and human-in-the-loop workflows
- ✓Researchers moving from NLP into human-AI interaction
Prerequisites
- •CS224N or CS229, or equivalent background in NLP or machine learning
- •Comfort reading research papers, since most lectures are built around assigned readings
FAQ
What is CS 329X: Human-Centered LLMs (Stanford, Fall 2026)?
Stanford's Fall 2026 graduate course on building LLM systems around the people who use them, taught by Diyi Yang. It is for engineers and researchers with an NLP or ML background who want the research behind RLHF, DPO, data curation, human-agent collaboration, interaction evaluation, and LLM safety and privacy, with public lecture slides and a full reading list.
Is CS 329X: Human-Centered LLMs (Stanford, Fall 2026) free?
CS 329X: Human-Centered LLMs (Stanford, Fall 2026) is free to access.
What level is CS 329X: Human-Centered LLMs (Stanford, Fall 2026) for?
CS 329X: Human-Centered LLMs (Stanford, Fall 2026) is aimed at a advanced audience. Recommended background: CS224N or CS229, or equivalent background in NLP or machine learning, Comfort reading research papers, since most lectures are built around assigned readings.
How long does CS 329X: Human-Centered LLMs (Stanford, Fall 2026) take?
Expect roughly 10-week quarter (Sep 23 to Dec 2, 2026), two lectures a week; public slides and readings, self-paced. Most learners work through it at their own pace.
What will I learn from CS 329X: Human-Centered LLMs (Stanford, Fall 2026)?
You'll learn: How RLHF and direct preference optimization turn human preferences into model behavior; How data choices such as LIMA-style curation and FineWeb filtering shape model quality; Methods for evaluating human-AI interaction beyond static benchmark accuracy scores; How to study and design human-agent collaboration, including the Collaborative Gym setup; Design approaches for full-stack human-AI systems and generative, elicitive user interfaces; How LLM adoption changes work, drawing on automation versus augmentation research; Risks in LLM use including privacy, safety, overreliance, sycophancy and AI companionship; Programming LLM pipelines with DSPy and comparing prompting techniques from a systematic survey.
Topics
Sources
This page was written from 3 sources, 1 on domains other than web.stanford.edu.