Large Language Models, Spring 2026 (ETH Zürich)
by ETH Zürich (Ryan Cotterell, Mrinmaya Sachan, Florian Tramèr)
ETH Zürich's LLM course: formal language-model theory, then fine-tuning, RAG, RLHF and LLM security, with public slides and course notes.
Overview
Large Language Models is ETH Zürich's spring 2026 course, taught by assistant professors Ryan Cotterell, Mrinmaya Sachan and Florian Tramèr with a team of fifteen teaching assistants. The lecture schedule runs in three blocks. Cotterell's group opens with modeling foundations: the formal definition of a language model, finite-state and recurrent models, the representational capacity of RNN and transformer LMs, tokenization, and decoding. Sachan's block moves to adaptation: transfer learning and fine-tuning, two sessions on parameter-efficient fine-tuning, prompting and in-context learning, multimodality, retrieval-augmented LMs, reinforcement learning for reasoning and inference-time compute, and instruction tuning with RLHF. The final block covers evaluation and benchmarks, then security: adversarial examples, watermarks, prompt injections, data poisoning, backdoors, model stealing, memorization, differential privacy and membership inference attacks, with Tramèr closing the privacy lectures. Weekly tutorials cover n-grams, RNN and transformer LMs, tokenization and generation, PEFT, RAG, chain-of-thought prompting, and watermarking, and some link exercises, solutions and notebooks. The theory half follows the group's course notes, published on arXiv as 'Formal Aspects of Language Modeling' (2311.04329). Many lecture slides are public on the course page. Recordings and assignments sit behind ETH's Moodle login, so outside learners get the slides, notes and tutorial material.
At a Glance
- Topic
- Models
- Level
- Advanced
- Format
- Course
- Cost
- Free
- Duration
- One semester (17 Feb to end of May 2026), ~3h lectures + 2h tutorial per week, self-paced from the public slides and notes
- Provider
- ETH Zürich (Ryan Cotterell, Mrinmaya Sachan, Florian Tramèr)
- Hands-on
- Yes — code/exercises
- Certificate
- None
What You’ll Learn
- ✓Define a language model formally and frame language modeling as a task
- ✓Compare the representational capacity of finite-state, RNN and transformer language models
- ✓Explain how tokenization and decoding choices change the text a model generates
- ✓Apply full fine-tuning and parameter-efficient fine-tuning methods to pretrained language models
- ✓Build retrieval-augmented language models and understand reinforcement learning for reasoning
- ✓Describe instruction tuning and RLHF, and how benchmarks evaluate the resulting models
- ✓Analyze prompt injection, data poisoning, backdoor and model stealing attacks on LLMs
- ✓Measure memorization and privacy leakage with differential privacy and membership inference
Highlights
- •Taught by three ETH faculty whose groups publish on LM theory, NLP and ML security
- •Unusually deep security block: five sessions on adversarial examples, prompt injection, poisoning, model stealing and privacy
- •Theory half is backed by the free 'Formal Aspects of Language Modeling' notes on arXiv
- •Public weekly tutorial topics with linked exercises, solutions and notebooks for several sessions
Who It’s For
Best For
- ✓ML engineers who want the formal theory behind transformers and decoding
- ✓Engineers responsible for LLM security, red teaming or privacy reviews
- ✓Graduate students preparing for NLP or LLM research
Prerequisites
- •Probability theory and linear algebra at university level
- •Machine learning fundamentals and some computational complexity
- •Comfort reading mathematical notation and proofs
FAQ
What is Large Language Models, Spring 2026 (ETH Zürich)?
A university course from ETH Zürich for engineers and graduate students who want the theory under large language models as well as the practice. It starts from what a language model formally is, works through RNNs and transformers, then covers fine-tuning, retrieval, RLHF, evaluation, and attacks such as prompt injection, data poisoning and membership inference.
Is Large Language Models, Spring 2026 (ETH Zürich) free?
Large Language Models, Spring 2026 (ETH Zürich) is free to access.
What level is Large Language Models, Spring 2026 (ETH Zürich) for?
Large Language Models, Spring 2026 (ETH Zürich) is aimed at a advanced audience. Recommended background: Probability theory and linear algebra at university level, Machine learning fundamentals and some computational complexity, Comfort reading mathematical notation and proofs.
How long does Large Language Models, Spring 2026 (ETH Zürich) take?
Expect roughly One semester (17 Feb to end of May 2026), ~3h lectures + 2h tutorial per week, self-paced from the public slides and notes. Most learners work through it at their own pace.
What will I learn from Large Language Models, Spring 2026 (ETH Zürich)?
You'll learn: Define a language model formally and frame language modeling as a task; Compare the representational capacity of finite-state, RNN and transformer language models; Explain how tokenization and decoding choices change the text a model generates; Apply full fine-tuning and parameter-efficient fine-tuning methods to pretrained language models; Build retrieval-augmented language models and understand reinforcement learning for reasoning; Describe instruction tuning and RLHF, and how benchmarks evaluate the resulting models; Analyze prompt injection, data poisoning, backdoor and model stealing attacks on LLMs; Measure memorization and privacy leakage with differential privacy and membership inference.
Topics
Sources
This page was written from 2 sources, 1 on domains other than rycolab.io.