Multimodal Foundation Models for Biomedicine:
Challenges and Opportunities

Technical Challenges

What are the core limitations of existing multimodal learning techniques when applied to biomedical data? Challenges include cross-modal alignment between data with different spatial and temporal resolutions; handling extreme data imbalances between well-annotated and sparse modalities; maintaining modality-specific contexts while enabling knowledge transfer across domains.

New Opportunities

What transformative opportunities do multimodal foundation models unlock in biomedicine? Potential breakthroughs include multi-scale disease diagnosis by combining radiology images with pathology slides; personalized treatment by integrating wearable sensor data with genomic profiles; context-aware operations by synchronizing surgical videos with patient records.

Challenges in Current Techniques

Papers highlighting limitations in existing methods, especially simple yet intuitive approaches that unexpectedly lead to negative outcomes.

• Multimodal foundation models - pre-training, post-training, alignment
• Agentic framework - design, evaluation, control
• Benchmark and evaluation - failure modes, reproducibility

Opportunities with MMFMs

Papers showcasing novel applications, especially in underexplored areas such as drug discovery and surgery.

Selected papers will be featured as posters, and four will receive spotlight oral presentations.