Imperial College · 2026
LLM & RL
Reading Group
A bi-weekly space to learn about reinforcement learning and its intersection with large language models, followed by friendly discussion.
Suggest a topicWhen
Every other Thursday
13:00
Where
Huxley 410
Imperial College London
Who
Researchers and students
with a basic ML background
Join
Contact Paul or
Sergio Estan Ruiz
The group
From foundations to current research.
We begin with core reinforcement-learning concepts—Markov decision processes, policy gradients, and value functions—and large-language-model fundamentals such as transformers, pretraining, and scaling laws.
From there, we move to contemporary topics including RLHF and reinforcement-learning approaches to reasoning and planning. Speakers explain a topic, then the room shapes the discussion around shared questions and interests.
Programme
Schedule
Select a session to read its abstract
Session 1 LLMs 101 Tom Coates
“What is this transformer thing, anyway?” I will describe what is going on inside a large language model, starting from fundamental definitions.
Session 2 LLMs 102 Tom Coates
Finishing the talk from session 1.
Session 3 Training and deploying LLMs Sara Veneziale
In this talk, we will go through a high-level overview of how modern LLMs are trained. We’ll highlight the different stages of training, focusing on the different ways that engineers can influence model behaviour at both training and inference stages, and pointing out where resource constraints—on either the data or compute side—become an issue.
Session 4 Training a small language model to speak fake Latin Paul Lezeau
A lab session: the goal is for everyone to get their hands dirty and implement what we have covered so far with a toy problem.
Resources
- Beginner notebook We’ll work through this together.
- Advanced notebook For participants already familiar with PyTorch.
- Solution notebook A working solution, open to improvement.
Session 5 Introduction to reinforcement learning Benjamin Sanati
In this talk, I will show how probability theory, statistical inference, and decision theory are used to describe agents that learn to act by interacting with their environment. After a philosophical overview, we will consider decision-making and learning under uncertainty, building towards the fundamental principles of reinforcement learning.
Session 6 Safety research in AI: are we screwed? Sergio Estan
I recently received a book called If Anyone Builds It, Everyone Dies: The Case Against AI Superintelligence. That seems a bit intense. After all, LLMs are just regression models trained on huge amounts of internet data, right? To what extent do new LLM-based systems pose a real threat—and what can we do to minimise such risks? This talk surveys the safety-research landscape around LLMs, including attacks on LLM systems and the defence mechanisms that researchers and practitioners use. Some of the talk draws on BlueDot Impact’s technical AI safety course.
Resources
- Technical AI Safety course BlueDot Impact