Flávia Carvalhido on Responsible Multimodal AI & Stress Testing

At the forefront of ensuring artificial intelligence serves humanity responsibly, Flávia Carvalhido, a PhD student at the University of Porto, is tackling a critical challenge: making multimodal AI models safer for high-stakes applications like medical diagnostics. Her research, part of the University of Porto’s Center for Responsible AI project, focuses on “Stress Testing of Image-Text Multimodal Models in Medical Image Report Generation,” a crucial area where AI’s precision directly impacts patient care. Supervised by Professors Henrique Lopes Cardoso and Vítor Cerqueira at the LIACC research laboratory, Carvalhido’s work aims to uncover the limitations of cutting-edge AI systems, fostering their more informed and reliable use in healthcare.

Carvalhido’s research is rooted in the understanding that for AI to be truly beneficial in safety-critical sectors, its vulnerabilities must be thoroughly understood. By subjecting state-of-the-art multimodal AI models—those that process both images and text, like generating a radiology report from a scan—to rigorous stress tests, she seeks to pinpoint their failure points. This methodology, still relatively underexplored in AI development, is intended to become a fundamental procedure in the creation of responsible AI. Having already defended her thesis proposal, Carvalhido is currently deep in the developmental phase of her PhD, which encompasses a broad literature review and a meticulously planned methodology spanning three years. Her work integrates several key disciplines: multimodal AI, medical image report generation, explainable AI, model robustness, responsible AI, and the novel application of stress testing.

One of the most compelling aspects of Carvalhido’s journey is navigating the uncharted territory of AI stress testing. This pioneering work offers unique opportunities for groundbreaking discoveries, yet it demands exceptional originality and rigorous, high-quality research to establish a new paradigm within the field. A significant challenge she faces is the sheer velocity of progress in multimodal generative AI. New research papers emerge weekly, necessitating constant adaptation and updating of her own work to remain relevant. This rapid evolution in AI development often outpaces the more deliberate pace required for responsible AI research, creating a delicate balance that is central to her PhD.

Looking ahead, Carvalhido plans to systematically explore various facets of medical image report generation. This involves an in-depth analysis and comparison of current leading models using Explainable AI (XAI) techniques to better understand their internal workings and performance. This foundational step will establish a representative baseline against which her initial stress testing methodologies can be applied. Her ultimate goal is to comprehend how these generative models behave under pressure. Following this, she will begin to probe the models with intentionally “challenging” or stress-inducing inputs, meticulously documenting any shifts in their performance or generative outputs. A crucial next phase will involve defining precisely what constitutes a stress-inducing input and, more broadly, solidifying the concept of stress testing within the AI domain.

Carvalhido’s path to AI research was sparked by a lifelong curiosity about how computers and the world operate. From an early age, a fascination with technology led her to pursue Informatics and Computer Engineering, where AI classes consistently captivated her. She is driven by the field’s boundless possibilities and its dynamic, ever-changing nature, believing AI will never cease to offer new learning and discovery. She aspires to leave a legacy of positive impact through her contributions to AI. For those considering a PhD in AI, Carvalhido emphasizes that it is both a profound challenge to one’s determination and a personal promise of commitment. She advises aspiring researchers to be prepared for constant change, remain adaptable, and stay laser-focused on their objectives, highlighting that the fast-paced nature of AI research demands proactive engagement and a willingness to dive in.

Her recent experience at the AAAI Doctoral Consortium and the broader AAAI conference proved transformative. As her first major conference, it offered an unparalleled opportunity to engage with diverse research ideas and perspectives on AI’s future. Meeting fellow young researchers and learning from seasoned professionals at such a large-scale event was an enriching experience that solidified her commitment to the field. Beyond her academic pursuits, Carvalhido, a 24-year-old from Porto, Portugal, is an avid traveler and a self-proclaimed foodie, having explored over ten European countries, often through volunteering. Her first journey outside Europe brought her to Philadelphia for AAAI 2025, and her dream destination remains Japan, drawn by the prospect of a profound cultural immersion.