Large language models in new AI research theme

CHAIR is enhancing Chalmers' expertise in AI by bringing together researchers from Chalmers and the University of Gothenburg around a few defined areas where AI is the central factor: the CHAIR themes. When the current themes were extended for an additional year, the theme Large Language Models for Mathematics and Programming was added.

Leverage of the creative process

The theme is set to explore how large language models, such as ChatGPT and similar, can be utilised to make new discoveries within mathematics and programming. The theme leaders believe that large language models can be used as a leverage for the creative process, of finding new and unexpected steps and proofs within both sciences.

At the same time, these models are known for making things up, or hallucinating, as it is called. Thus, there is also a need for simultaneous proof-checking of what the model suggests. The theme's focus is therefore to promote efficient combinations of large language models and logical methods such as symbolic models, and software such as proof assistants, that can ensure mathematical correctness.

"The theme is dedicated to exploring the potentials of combining these methods and what it can bring to mathematics and programming, leading to new tools for mathematicians and computer scientists, to accelerate research and development. We are excited to see what happens when we bring intuition and associative functions into these rule-based systems," says Moa Johansson, also theme leader.

Beyond the known rules

Today's large language models are performing very well when it comes to tasks such as building software or solving mathematical problems. As long as the language model knows the rules, it can be a very efficient tool, but Johan Wästlund sees potential of reaching beyond the known rules.

"ChatGPT and similar large language models are already able to solve complex mathematical problems in just a blink of an eye. They have learned the calculations from already existing program code. Now imagine AI that can reason mathematically and is able to create its own new methods that we never even thought of. For me, as a mathematician, this perspective is mind-blowing," he says.

The theme starts in August with seminars, workshops, and other events to gather Chalmers researchers that are interested in this field and to create a strong community.

The other themes will continue

Because of the success of the other four CHAIR themes - which have been active for over a year - and their work in community building within different fields, they have all been extended for yet another year. The themes are AI for Scientific Data Analysis, Interpretable AI, Structured Learning, and AI and Robotics for Health and Rehabilitation.

Read about all the themes and their activities at CHAIR's website and in the CHAIR calendar.

Updated 27 August 2024, 10:38Published 27 August 2024, 08:30

Leave feedback