How Do Language Models Reason? Foundations, Limitations and Neuro-Symbolic Approaches

Speaker: André Freitas (Idiap Research Institute & University of Manchester)
Series: Language Technology Lab Seminars
Also open online: the joining details are on the talk's page.

How do language models reason, and what would it take to build AI systems capable of sustained, complex reasoning? This lecture develops the foundations for answering these questions, connecting what we know about the internal workings of language models with methods for evaluating and extending their capabilities.

We begin by examining how language models represent knowledge and how training and inference shape their ability to solve problems. We explore what behavioural experiments and the investigation of internal mechanisms can reveal about these abilities, and where our understanding remains incomplete. We then consider o methods to qualify the spectrum of reasoning problems in a rigorous manner. Building on these foundations, we examine how to develop systems that can tackle reasoning tasks extending beyond the capabilities of an individual model. Neuro-symbolic and agentic approaches provide ways to connect language models with explicit knowledge, memory, computational tools and verification mechanisms. We investigate how these components can support the decomposition of problems, exploration of hypotheses, integration of evidence and revision of conclusions. Examples from mathematics and biomedical research will illustrate the opportunities and challenges of coordinating these processes, including maintaining coherence across many steps, managing uncertainty and evaluating whether each component contributes to a more reliable solution.

Bio: André Freitas leads the Neuro-Symbolic AI group at the Idiap Research Institute (Switzerland) and holds an honorary appointment at the University of Manchester (UK). His research investigates how AI systems reason and how to make that reasoning more reliable, explainable and verifiable. He develops neuro-symbolic and agentic approaches that connect language models with explicit knowledge, formal methods and computational tools to address complex problems in mathematics and scientific discovery, having published over 120 papers at the main conference s and journals in the NLP/AI space.