Master’s Thesis: Deconstructing Reasoning Traces in LLMs

AI Sweden•Gothenburg, NE

2d•Hybrid

About The Position

As Sweden's national center for applied AI, AI Sweden is seeking an ambitious Master’s thesis student to join their team. This project, conducted in close collaboration with the Stef SM Lab at Chalmers University of Technology, focuses on deconstructing reasoning traces in Large Language Models (LLMs). Modern AI systems utilize complex reasoning processes, but understanding the nature of these traces is crucial for reliability, security, and auditing. The project will explore topics such as adversarial robustness, honesty of reasoning, distinguishing reasoning from recall, and the visibility of cognitive processes in LLMs. Students will design and conduct controlled experiments using open-weight reasoning models, with access to substantial compute resources. The goal is to contribute to a frontier area of AI research, with potential for publication.

Requirements

Ongoing Master’s studies in Computer Science, Data Science, Engineering Physics, Complex Adaptive Systems, Machine Learning, or similar.
Comfortable with Python and deep learning frameworks.
Comfortable with the idea that an experiment might yield unexpected results.
Genuine curiosity and interest in AI safety and security.

Nice To Haves

Prior experience with language models.
Prior experience with reinforcement learning.

Responsibilities

Design controlled experiments to understand the reasoning phenomenon in LLMs.
Run experiments on open-weight reasoning models.
Measure and analyze experimental results carefully.
Collaborate with AI scientists and researchers at AI Sweden and Chalmers University of Technology.
Potentially contribute to publications at competitive venues.