Reasoning Under Uncertainty: Trustworthy LLMs for High-Stakes Medicine
Event details
| Date | 26.06.2026 |
| Hour | 09:30 › 11:30 |
| Speaker | Yusuf Kesmen |
| Location | |
| Category | Conferences - Seminars |
EDIC candidacy exam
Exam president: Prof. Robert West
Thesis advisor: Prof. Mary-Anne Hartley
Co-examiner: Prof. Martin Jaggi
Abstract
Current machine learning systems are very good at tasks reducible to statistical pattern recognition but falter on problems requiring deliberate, structured inference. Clinical reasoning exemplifies the latter: physicians generate hypotheses abductively, revise beliefs probabilistically as evidence accumulates, and choose among actions whose costs are deeply asymmetric. The harm of a missed diagnosis is not the harm of an unnecessary test. Large language models, increasingly deployed for clinical decision support, inherit this limitation. They conflate knowledge with inference: the same parameters that store medical facts are asked to reason over them, which leaves no explicit mechanism to quantify uncertainty, to abstain when evidence is insufficient, or to weight outcomes by their true clinical costs. This work argues that progress toward trustworthy medical AI requires separating the representation of knowledge from the machinery of inference, and grounding the latter in formal frameworks drawn from statistical decision theory. We investigate how these frameworks can be embedded into LLM reasoning pipelines and imposed as external scaffolding around them, maintaining competing hypotheses, guiding evidence gathering, and selecting actions that reflect the asymmetric cost structure of clinical practice. The goal is a class of systems capable of reasoning over their uncertainty, systems that know when they do not know.
Selected papers
coming soon
Exam president: Prof. Robert West
Thesis advisor: Prof. Mary-Anne Hartley
Co-examiner: Prof. Martin Jaggi
Abstract
Current machine learning systems are very good at tasks reducible to statistical pattern recognition but falter on problems requiring deliberate, structured inference. Clinical reasoning exemplifies the latter: physicians generate hypotheses abductively, revise beliefs probabilistically as evidence accumulates, and choose among actions whose costs are deeply asymmetric. The harm of a missed diagnosis is not the harm of an unnecessary test. Large language models, increasingly deployed for clinical decision support, inherit this limitation. They conflate knowledge with inference: the same parameters that store medical facts are asked to reason over them, which leaves no explicit mechanism to quantify uncertainty, to abstain when evidence is insufficient, or to weight outcomes by their true clinical costs. This work argues that progress toward trustworthy medical AI requires separating the representation of knowledge from the machinery of inference, and grounding the latter in formal frameworks drawn from statistical decision theory. We investigate how these frameworks can be embedded into LLM reasoning pipelines and imposed as external scaffolding around them, maintaining competing hypotheses, guiding evidence gathering, and selecting actions that reflect the asymmetric cost structure of clinical practice. The goal is a class of systems capable of reasoning over their uncertainty, systems that know when they do not know.
Selected papers
coming soon
Practical information
- General public
- Free