Universidade Nova de Lisboa (Portugal) - Press Release: The research proposal on how to make Large Language Models (LLMs) more reliable, by David Semedo and João Magalhães, professors in the Department of Informatics at NOVA FCT and researchers in the Multimodal Systems group at NOVA LINCS, has been selected for the Amazon Nova AI Challenge.

The team, which also includes students Diogo Silva (team leader), Rafael Ferreira, and Diogo Tavares, has gained new and important additions: João Pedro Soares, Daniel Pina, Iago Paulo, Artur Horal, and Henrique Paz. Once again, they are the only Portuguese team in this international competition.

This edition focuses on the theme "Trustworthy AI: Advancing AI-Assisted Software Development Safely to Create More Secure and Reliable Applications." Building on previous challenges—where innovative LLMs were developed for task assistance and fundamental security issues were addressed—this challenge aims to achieve security by design in Generative Artificial Intelligence. The research challenge focuses on generating secure code with LLMs that are robust against attacks, involving the participation of 10 universities worldwide, including Carnegie Mellon University, University of Illinois Urbana-Champaign, Czech Tech University, University of California at Davis, Virginia Tech, University of Texas at Dallas, Columbia University, University of Wisconsin, Purdue University, and NOVA FCT.

The NOVA FCT team will investigate methods to test the limits of LLMs and break their behavior to expose vulnerabilities and identify ways to improve their reliability. In 2023, the team developed the best LLM-based system for assisting in the execution of manual tasks.