Agentic Large Language Models

A master course (6 EC) at LIACS on large language models that reason, act and interact. We follow the textbook Agentic Large Language Models (Van Duijn, Van der Meer, Plaat & Van Stein) and teach the course together as the same four authors.

The course website, with the syllabus, the weekly schedule and announcements, is at agenticllms.liacs.nl.

What the course covers

The course follows the parts of the book:

• Language modelling: the foundations of LLMs, the training pipeline, scaling laws and efficient models.
• Reasoning: reasoning models; vision, action and world models; metalearning.
• Agents: interaction styles for agentic LLMs.
• Tools and applications: search and retrieval, how agents use tools to act in the world, and real-world assistants.
• Behavioural and cognitive perspectives: studying the behaviour of agentic LLMs, role-based interaction in multi-LLM systems, and emergent behaviour in multi-agent environments.
• Advanced topics: synthetic data generation, mechanistic interpretability for LLM agents, and cognitive perspectives.

How it works

There is no written exam. Students work on four hands-on Python case studies: building an LLM from scratch, reasoning and self-reflection with the BALROG benchmark, a multi-agent diagnosis framework, and simulating societies of LLM agents. The final grade is the average of the four.

Students are expected to have a background in Python and PyTorch, classical machine learning, deep learning, reinforcement learning and NLP basics; the first part of the course reviews the fundamentals.

A course that moves with the field

Agentic LLMs change from one semester to the next. We plan to update the book every year, and to update the course with it.