Postdoctoral Researcher in Multimodal Reasoning Models for Oncology

ETH Zurich seeks a Postdoctoral Researcher to develop and adapt multimodal reasoning models for oncology, focusing on clinical decision support through integration of patient data, literature, guidelines, and molecular evidence.

This role focuses on advancing foundation models tailored for oncology, enabling them to reason across complex clinical scenarios such as diagnosis, treatment planning, and longitudinal patient care by integrating diverse data sources and ensuring safety, reliability, and traceability.

Responsibilities

• Design and implement multimodal language model architectures for oncology applications.
• Integrate clinical context, biomedical literature, treatment guidelines, and patient-level multimodal data into model systems.
• Adapt and evaluate models using public and institutional oncology datasets.
• Develop reasoning behaviors that account for uncertainty and safety in clinical decision-making.
• Create model workflows that reliably interact with external tools and knowledge sources in an auditable manner.
• Retrieve information from clinical literature, guidelines, and trial databases to support decision-making.
• Implement clinical trial matching and therapy evidence lookup functionalities.
• Incorporate variant interpretation and molecular knowledgebase use into model pipelines.
• Build multi-agent systems to break down complex oncology tasks into structured, hierarchical reasoning streams.
• Generate citation-grounded, traceable outputs suitable for expert review.
• Develop post-training methods to enhance clinical reasoning quality, reliability, and safety.
• Apply process-level supervision to monitor intermediate reasoning steps.
• Use outcome-based supervision with expert or guideline-derived signals.
• Apply reinforcement learning techniques to shape oncology-specific reasoning behavior.
• Compare and refine reinforcement learning training approaches.
• Implement calibration, abstention, and safety-aware optimization strategies.
• Evaluate models in clinically meaningful contexts, including guideline concordance, diagnostic and therapeutic reasoning quality, molecular interpretation accuracy, tool-use reliability, citation quality, and evidence grounding.
• Ensure calibration, uncertainty estimation, and appropriate deferral in model outputs.
• Support trace auditability and clinician-in-the-loop evaluation methods.