Postdoctoral Researcher in Multimodal Reasoning Models for Oncology

100%, Basel, fixed-term

We are seeking an exceptional and highly motivated Postdoctoral Researcher to lead research on multimodal reasoning models for oncology. The project focuses on developing, post-training, and evaluating flexible AI models that can support complex oncologic diagnostic and therapeutic decision-making in a safe, transparent, and clinically grounded manner.

The successful candidate will work on oncology-focused multimodal reasoning models that combine language, vision, biomedical knowledge, clinical context, and relevant patient-level data to produce reliable, auditable, and uncertainty-aware outputs.

A major focus of the position is the development of AI-based reasoning strategies for oncology, including tool-augmented inference, multi-agent or compound model workflows, process supervision, verifier-guided training, and reinforcement learning-based post-training. The goal is to build systems that can justify recommendations, cite supporting evidence, calibrate uncertainty, defer appropriately, and operate safely in clinically realistic settings.

This position is embedded within a highly interdisciplinary collaboration between ETH Zurich, Kaiko.ai, and clinical partners, offering an opportunity to advance foundational AI research while working toward real-world translation in oncology.

Job description

Reasoning Models for Oncology

Development and adaptation of oncology-focused foundation models capable of reasoning over complex clinical questions, including diagnosis, molecular interpretation, treatment selection, and longitudinal care.

This may include:

• Multimodal language model architectures
• Integration of clinical context, biomedical literature, guidelines, and patient-level multimodal evidence
• Adaptation and evaluation on public and institutional oncology datasets
• Development of uncertainty-aware and safety-aware reasoning behavior

Reasoning Strategies, Agents, and Tool Use

Development of model workflows that can use external tools and knowledge sources in a reliable and auditable way.

Examples include:

• Retrieval from literature, clinical guidelines, and trial databases
• Clinical trial matching and therapy evidence lookup
• Variant interpretation and molecular knowledgebase use
• Multi-agent systems for decomposing complex oncology tasks into hierarchical context streams
• Citation-grounded and traceable outputs suitable for expert review

Process Supervision and Post-Training

Development of post-training methods that improve clinical reasoning quality, reliability, and safety.

This may include:

• Process-level supervision for intermediate reasoning steps
• Outcome-based supervision using expert or guideline-derived signals
• Reinforcement learning for oncology-specific reasoning behavior
• Comparison and development of RL training approaches
• Calibration, abstention, and safety-aware optimization

Clinical Evaluation and Safety

Evaluation of oncology reasoning models in clinically meaningful settings.

Key evaluation dimensions include:

• Guideline concordance
• Diagnostic and therapeutic reasoning quality
• Molecular interpretation accuracy
• Tool-use reliability
• Citation quality and evidence grounding
• Calibration, uncertainty, and appropriate deferral
• Trace auditability and clinician-in-the-loop evaluation Profile

Must Have

• PhD in Computer Science, Machine Learning, Medical AI, Biomedical Informatics, Computational Biology, or a related field
• Strong programming skills in Python and modern ML frameworks
• Experience with deep learning and large language models
• Strong publication record in AI/ML, medical AI, computational biology, biomedical informatics, or related areas
• Ability to work in highly interdisciplina