CogSci seminar - 26th November
When: Thursday, November 26, 2026, 12:00 PM - 1:30 PM
Where: Queens Building, LG2, Mile End campus
We are pleased to invite you to an additional edition of the CogSci seminar next week: a Regensburg Special featuring Prof. Udo Kruschwitz (University of Regensburg, and long-term collaborator of Massimo Poesio) alongside his two PhD students, Gregor Donabauer and Markus Bink.
The talk will take place next Thursday, 26th of November, from 12:00 – 13:30 at Queens Building, LG2. And can also be joined via MS Teams; use this link: https://teams.microsoft.com/meet/372892441300302?p=kSHeIleCxtlY0O8q92
Talk 1 Speaker:
Gregor Donabauer
Title:
Opportunities and Challenges of LLMs in Healthcare: Insights from Two Case Studies
Abstract:
The recent advances in Natural Language Processing (NLP) and Large Language Models (LLMs) present opportunities across demanding medical scenarios, including automated treatment recommendation extraction, clinical guideline harmonisation, and ultimately, clinical decision-making. However, applying NLP in healthcare remains particularly challenging since clinical decision-making is a high-stakes domain where recommendations need to be consistent and reliable, and factual errors can directly affect patient safety and health-related outcomes. Standard LLM workflows face various challenges, including model non-determinism, which can lead to output variations across repeated runs, as well as the complexity of long clinical guideline documents that may contain conflicting treatment recommendations and multimodal decision pathways.
In this talk, we share insights from our work on two methodological workflows designed to address these challenges by systematically analysing and harmonising medical evidence. First, we demonstrate how model non-determinism can be mitigated through repeated runs and semantic clustering, resulting in more robust extraction of information nuggets. We then show how multimodal treatment recommendation extraction from clinical guidelines, combined with cross-document stance harmonisation, can help to identify divergences between guidelines and evidence gaps in the literature.
Finally, we provide an outlook on an ongoing project that deploys open-weight LLMs locally within the hospital to generate guideline-grounded treatment recommendations for multidisciplinary tumour boards. Local deployment is essential for protecting sensitive patient data, however, processing clinical guidelines with more than 500 pages with local models that have limited context windows presents another major challenge. We conclude with a preview of how Retrieval-Augmented Generation (RAG) can help to address these context limitations and support guideline-grounded clinical decision-making.
Talk 2 Speaker:
Markus Bink
Title:
From Answer Engines to Learning Companions: Rethinking Support in Information Seeking
Abstract:
People increasingly turn to search engines and AI assistants for answers to important questions, from health decisions to the claims they come across online. Generative AI has made this easier than ever. Instead of a list of links to read and compare, users now get a single fluent, confident answer. That convenience comes at a cost. Polished answers invite overreliance, discourage users from checking sources themselves, and can quietly pass along errors and biases. Many people also overestimate their own ability to judge information online, which makes these risks harder to notice. As a result, the very skills needed to navigate today’s information landscape may be weakening just when they matter most.
In this talk, we make the case for pedagogical search systems: tools that teach people to search and evaluate information while they use them. Drawing on two user studies in health information seeking, we compare two different approaches. One is a Socratic-inspired copilot that coaches users instead of answering them. The other is a lightweight companion that offers short, well-timed tips during search. We showcase the potential and pitfalls of such pedagogical support, and outline avenues for future work. These include adapting support to users’ needs and expertise, timing guidance to the moment it is useful, and using friction deliberately rather than by default.
Key contact: Juntao Yu