CMSC 25750/35750: Large Language Models (UChicago, Spring 2026)
by University of Chicago (Chenhao Tan)
A paper-by-paper reading course on LLM interpretability, alignment failures and agents from UChicago's Chenhao Tan.
Overview
CMSC 25750/35750 is the spring 2026 Large Language Models course at the University of Chicago, taught by Chenhao Tan, an associate professor of computer science and data science who directs the Chicago Human+AI lab. It is a research course, not a build-along tutorial: each session pairs a topic with two or three required papers, and grading leans on a research project (50%), quizzes, assignments and weekly blog entries. The schedule has three parts. Interpretability covers attention analysis, MLPs and factual recall (ROME, knowledge neurons), transformer circuits and induction heads, the logit lens, the geometry of representations, superposition and sparse autoencoders, chain-of-thought faithfulness and monitorability, and interpretability applied to protein language models. Alignment covers the alignment problem, scalable oversight through debate and weak-judges-strong setups, emergent misalignment, sycophancy and the limits of RLHF, sleeper agents, sandbagging and in-context scheming. Agents, with a guest lecture from SWE-bench author Ofir Press, covers SWE-bench, SWE-agent, Agentless and ReAct, then research agents such as AlphaEvolve and Agent Laboratory, LLM simulation of people, and complementary human-AI decision making. Readings run through 2026. The course requires prior NLP and generative-model coursework and some experience training models.
At a Glance
- Topic
- Models
- Level
- Advanced
- Format
- Course
- Cost
- Free
- Duration
- 10-week quarter (23 Mar to 22 May 2026), 17 lecture sessions with ~3 papers each; ~40-60 hours to read the full list
- Provider
- University of Chicago (Chenhao Tan)
- Hands-on
- No
- Certificate
- None
What You’ll Learn
- ✓Trace how transformer circuits and induction heads implement in-context learning
- ✓Use sparse autoencoders and superposition results to interpret model features
- ✓Judge whether a chain of thought faithfully reflects what the model computed
- ✓Explain scalable oversight methods such as debate and weak models judging strong ones
- ✓Recognize emergent misalignment, sycophancy, sandbagging and scheming in model behavior
- ✓Evaluate coding agents with SWE-bench and compare SWE-agent with Agentless designs
- ✓Assess research agents like AlphaEvolve and LLM simulations of human survey data
Highlights
- •A curated, ordered reading list of roughly fifty papers spanning 2016 to 2026
- •Agents session guest-lectured by Ofir Press, co-author of SWE-bench and SWE-agent
- •Covers recent safety results such as emergent misalignment and chain-of-thought monitorability
- •Rare interpretability-for-science session on sparse autoencoders for protein language models
Who It’s For
Best For
- ✓Engineers moving into interpretability or alignment research
- ✓Agent builders who want the research behind coding-agent benchmarks
- ✓Graduate students looking for a structured LLM research reading list
Prerequisites
- •Natural language processing coursework (CMSC 25700/35100 or equivalent)
- •Generative models coursework and a working understanding of the transformer architecture
- •Experience training and analyzing language models; research experience preferred
FAQ
What is CMSC 25750/35750: Large Language Models (UChicago, Spring 2026)?
A research-oriented University of Chicago course whose public schedule is a curated reading list of about fifty papers. It suits engineers and researchers who already know transformers and want to understand how LLMs work inside, how alignment breaks, and how coding and research agents are built and evaluated.
Is CMSC 25750/35750: Large Language Models (UChicago, Spring 2026) free?
CMSC 25750/35750: Large Language Models (UChicago, Spring 2026) is free to access.
What level is CMSC 25750/35750: Large Language Models (UChicago, Spring 2026) for?
CMSC 25750/35750: Large Language Models (UChicago, Spring 2026) is aimed at a advanced audience. Recommended background: Natural language processing coursework (CMSC 25700/35100 or equivalent), Generative models coursework and a working understanding of the transformer architecture, Experience training and analyzing language models; research experience preferred.
How long does CMSC 25750/35750: Large Language Models (UChicago, Spring 2026) take?
Expect roughly 10-week quarter (23 Mar to 22 May 2026), 17 lecture sessions with ~3 papers each; ~40-60 hours to read the full list. Most learners work through it at their own pace.
What will I learn from CMSC 25750/35750: Large Language Models (UChicago, Spring 2026)?
You'll learn: Trace how transformer circuits and induction heads implement in-context learning; Use sparse autoencoders and superposition results to interpret model features; Judge whether a chain of thought faithfully reflects what the model computed; Explain scalable oversight methods such as debate and weak models judging strong ones; Recognize emergent misalignment, sycophancy, sandbagging and scheming in model behavior; Evaluate coding agents with SWE-bench and compare SWE-agent with Agentless designs; Assess research agents like AlphaEvolve and LLM simulations of human survey data.
Topics
Sources
This page was written from 2 sources, 1 on domains other than uchicago-llm-course.github.io.