Science Research in Industry Experience Programme (SRIE)

Empowering undergraduate mathematicians through industry research.

About SRIE

The Science Research in Industry Experience (SRIE) is a mentored research programme designed specifically for undergraduate mathematics students.

The team behind SRIE consists of a network of mathematics alumni from Cambridge and other universities, united by the desire to help mathematics students get acquainted with pressing science issues in industry.

Through real-world research projects designed and led by industry researchers, SRIE hopes to help students discover how they can make unique contributions based on their rigorous academic training, and accelerate their careers to leadership positions within their field of choice.

Research Streams

Deep-dive industry research led by founders and fellowship scholars for undergraduate mathematicians.

The Gap Attractor (In Collaboration with Sentient Futures)

Julia Bossmann

MENTORS

OVERVIEW

The Gap Attractor investigates whether frontier language models exhibit stable, self-directed behavioural patterns when given freedom to explore without an explicit task or reward.

The project will produce a documented experimental battery and reproducible harness, a dataset of model trajectories and coded behavioural results, and a co-authored research preprint with citable supplementary materials.

Intelligent Agency

Ashe Vazquez Nuñez

MENTORS

OVERVIEW

This stream is broadly dedicated to the development of a formal, rigorous theory of intelligent agency. Such a theory would unify perception and action to understand beings that exchange information and resources with their environment. These theories would aim at explaining AI and human agency, as well as how these interact. Experiments on AI cognition that relate to conceptual work are well within scope.

First-Order Definability of Policies

MENTORS

Vardhan Kumar Ray

OVERVIEW

Which agent functions can be written in first-order logic depends on three choices: how much structure
a single observation carries, how much memory the policy has, and which logic is meant. Fixing them
yields a small grid of questions. Three of the cells are closed by existing theorems, recorded here; the rest
are stated as three open problems.

World Models

Jennifer Benedict

MENTORS

OVERVIEW

World models matter for agent foundations because an entity needs one to optimise or act at all. Some worlds seem to admit more natural models than others (e.g., Newtonian mechanics is a very natural way to model ours). How close an AI's world model is to ours affects both how predictable its actions are and how similar its values are to our own. It's unclear what type of thing a world model should be, partly because it's unclear what type of thing the world is. Models may have a different type than the world, and will certainly be lossy or coarse-grained relative to it.

Learn More