Topic: Who decides what's trustworthy? Standards development for the EU AI Act

When: 27 August at 12.00-13.00

Where: Online - link by registration

Speakers:

Andreas Bexell, Industrial PhD student, Software Development and Environments, Lund University and Ericsson

Spoken language: English

Abstract

The use of large language models is sweeping into many fields and is being explored for many natural language tasks, ranging from text generation to classification. The introduction of tools like ChatGPT and co-pilot has been transformative for software development, engineering, and education. Many efforts aim to employ natural language processing and AI for different applications and domains throughout the public and private sectors.

Large language models provide impressive results but require more time and energy than other models. For some tasks, simpler models may be more sustainable.

In an experiment, we construct natural language classifiers of different complexity for a use case in the software engineering domain and demonstrate results similar to the state-of-the-art.

We find that simpler models perform similarly to LLMs while using a fraction of the energy. We also find that contrary to expectations, the classifiers based on large models do not generalize better to previously unseen data. Based on our results, we recommend careful model selection and that models for practical applications should be specialized using training data from the target set. Choosing a small model can lead to a significant reduction in power usage without compromising performance.

Registration

To participate is free of charge. 
Sign up at ai.lu.se/2025-08-27/registration and we send you an access link to the zoom platform.