Master’s Thesis: Deconstructing Reasoning Traces in LLMs

As Sweden's national center for applied AI, we're on a mission to accelerate the use of AI to benefit our society, our

competitiveness, and everyone living in Sweden. We drive impactful initiatives in areas such as healthcare, energy, and public

services while pushing the boundaries of AI research in fields such as natural language processing, machine learning and AI

security. Join us in harnessing the untapped value of AI to drive innovation and create sustainable value for Sweden.

We are now looking for an ambitious Master’s thesis student to join us at AI Sweden, in close collaboration with the Stef SM Lab

at Chalmers University of Technology.

Modern AI systems write out long chains of reasoning, checking their own work, backtracking when something looks wrong. A lot of

the recent progress in AI capabilities has come from this, and most agentic systems now depend on it. But we understand

surprisingly little about what these reasoning traces actually are. Is the written reasoning the computation, or a story told

about a computation that happened elsewhere? Does a model that reasons longer become more reliable, or merely more confident? What

are the parameters of reasoning as an attack surface against models? Reasoning traces are becoming a security control: if we can

read a system's reasoning, we can audit it. That only works if the trace is honest, robust, and not something an adversary can

steer.

A project may focus on one of the following topics or combine several of them. The exact research question and methods will be

decided together with the student, depending on the student's background and interests, previous work, and practical feasibility:

• Adversarial robustness for reasoning traces. Small changes injected into a model's reasoning pathway can have long-lasting,

devastating effects. Can the model recover, or does it absorb the error and build on it? Does thinking longer help or hurt?

• Honesty. When a model backtracks on its own logic or decisions, does anything actually change inside it, or is the correction

after the fact? Do reasoning traces define the intent, or are they just subvertible representations of it?

• Reasoning or recall? Is reasoning a recall mechanism (eliciting solutions that the model somehow already knows), or does it

contribute to unlocking a fundamentally new class of solutions that would not be attainable otherwise?

• Visibility. Do reasoning traces reveal the full cognitive process, or are there other mechanisms behind the scenes that models

rely on?

Projects can be theoretical, computational, or a combination of the two. The main goal is to use a simplified model or controlled

experiment to understand the reasoning phenomenon: Students will design controlled experiments, run them on open-weight reasoning

models, and measure carefully. You will have access to substantial compute, and you will be working on questions that are

genuinely open, in a frontier area of AI research, so there is a real chance your results are the first anyone has. The thesis can

be done alone or in pairs.

• Ongoing Master’s studies in Computer Science, Data Science, Engineering Physics, Complex Adaptive Systems, Machine Learning, or

similar.

• You should be comfortable with Python and deep learning frameworks, and comfortable with the idea that an experiment might tell

you something you did not expect.

• Prior experience with language models and reinforcement learning is a plus, but what matters most is a genuine curiosity and

interest in AI safety and security.

At AI Sweden, we are committed to building diverse and inclusive teams. Some positions may be subject to export control

regulations, which means that specific requirements may apply.

Doing your thesis at AI Sweden means working alongside leading AI scientists, change leaders, journalists, linguists, policy

professionals and entrepreneurs—all working with a higher purpose in mind than “just” tech. AI Sweden is Sweden’s National Center

for AI, we drive research questions that have both a long shelf-life and are widely applicable to Swedish industry and the public

sector. We aim for publications at the most competitive venues and celebrate a culture of research excellence.

As an organization, we’re uniquely positioned at the sweet spot of governmental influence and startup agility. Small enough to

stay adaptive and have fun but backed by and in close contact with both the government, academia and private and public sector. We

offer our team members a place to grow, an environment for personal development and achievements.

This Master’s thesis will be carried out at AI Sweden at Lindholmen Science Park, Gothenburg, in close collaboration with the Stef

SM Lab (https://stefsmlab.github.io/) at Chalmers. We aim for a hybrid working mode, with some time spent on-site, but there is

room for flexibility.

Application Deadline: 2026-12-27, (rolling selection – position may be filled earlier).

Start Date: January 2027

Contact

Mauricio Muñoz, Senior Research Engineer and Project Lead at AI Sweden

AI Sweden does not accept unsolicited support and kindly ask not to be contacted by any advertisement agents, recruitment agencies

or manning companies.