Artificial Intelligence (AI) models are now capable of generating text, answering questions in seconds and solving complex problems across multiple domains, but “how does a machine learn to make decisions?”. This question set the tone for the lecture “Reinforcement Learning in the Era of Large Language Models: Challenges and Opportunities”, by Zita Marinho, a professor at Instituto Superior Técnico, Universidade de Lisboa, held as part of the Tech Conference (SINFO) at Técnico, from 20 to 24 April at Técnico Innovation Center powered by Fidelidade.
The Técnico professor explored various Reinforcement Learning (RL) methods for training large language models. Among examples of systems trained with human feedback and automatic response verification mechanisms, she highlighted the potential of these technologies to support scientific discovery and solve complex problems.
“We are trying to build systems that do not merely generate plausible text, but that can learn to solve tasks in a more robust and useful way”, she explained. “The big challenge is figuring out how to create models that can ‘reason’ better, adapt to new contexts and produce reliable answers”.
The professor, a former Senior Research Scientist at Google DeepMind – a technology company focused on developing large-scale machine learning systems – highlighted computational resources as one of the main obstacles to the development of these models. “Current models require enormous processing power and large volumes of data”, she said. “A significant part of the research involves finding more efficient methods that can accomplish more with fewer resources”.
For Zita Marinho, this “directly influences the future of academic research”. “If we want to democratise the development of AI, we need methods that do not rely exclusively on massive infrastructure”, she added, emphasising the importance of reducing barriers to “access computing power” and promoting more efficient and sustainable approaches. “Innovation also involves making these technologies more accessible to the scientific community”.
The conversation with the students went beyond the technical aspects of the technology. When asked about academic pathways and opportunities in the field, the professor highlighted the importance of encouraging more young people to get involved in AI research. “These initiatives help students understand the options available at Técnico and the areas where they can work”, she said. “Creating a strong Técnico’s Artificial Intelligence community also means creating more opportunities for students”.
“Students stop just reading about innovation and start discussing it with those who create it”
“SINFO serves as a guide to the latest trends in the technology sector”, argues Cecília Correia, a Técnico alumna (Computer Science and Engineering) and coordinator of this year’s event. “By presenting the latest trends, we help students align their training with market demands”. For her, one of the central aspects of this initiative is the direct interaction between students, companies and researchers. “Students stop just reading about innovation and start discussing it with those who create it”.
“We want to highlight these realities through the successful careers of our guest speakers”, says Cecília Correia, emphasising the variety of career paths represented throughout the event. “It is common for former SINFO organisers to return as representatives of their own companies”, she noted. “This culture of continuity and excellence defines us”.
The 33rd edition of SINFO is part of the Técnico Career Weeks and brought together over 6,000 participants, around 100 companies and dozens of national and international speakers. Through workshops, lectures and networking opportunities, the initiative once again transformed Técnico into a hub for education, research and the technology industry, bringing students closer to the main trends and current challenges in Computer Science.
