Course overview: Understanding LMs#

This course deals with language models (LMs), in particular (but not exclusively so) on transformer-based language models like GPT-x or LLama. It covers topics that will equip participants with a better conceptual and practical understanding of what LMs are, how they work, and how to understand them. In the course, we will look (among others) at LM architectures, training & fine-tuning, prompting, mechanistic interpretability of LMs, LM agents and different evaluation methods. Participants will be offered both a technical perspective and encouraged to critically think about important topics relevant to cognitive science and society in the context of LMs.

Intended audience#

The course is intended for master students or advanced bachelor students (of, e.g., (computational) linguistics, cognitive science, or computer science) who are interested in language models / NLP / linguistics / AI / interdisciplinary approaches to state-of-the-art technical advances. Prior knowledge of LMs is not required; prior experience with programming in Python is highly encouraged.

Course formalia#

In SS 2025, the course consists of a weekly lecture (Tue, 9-12) and a weekly practical seminar (Thu, 14-16) which will systematically introduce and cover concepts from the lecture hands-on, in practical worksheets, where participants will work with (small) LMs themselves. This webbbook will host both lecture materials and practical materials.

The course is intended for 6 ECTS. There will be compulsory homework assignments and a final exam. Optionally, for 9 ECTS, students will conduct group projects.

Schedule#

The lecture overview below is preliminary and subject to changes.

  1. Introduction

  2. Neural networks and small language models

  3. Transformers

  4. Large language models, prompting & fine-tuning

  5. Probing & attribution

  6. Evaluation

  7. Mechanistic interpretability

Further materials#

There are various courses on deep learning, NLP and LMs out there. To name a few, these are excellent sources for further related materials:

Andrej Karpathy has an excellent 1h overview video.

We also highly recommend Brian Christian’s book The Alignment Problem for a more general perspective on AI and computational CogSci.

Previous iterations of the course#

Slides from the 2024 iteration of the course can be found here.