Neuro-symbolic foundation models (T1)

The main objective is to enable compositional reasoning across multiple modalities: language, vision, and structured knowledge. The dual aim is to reduce the “reasoning gap” caused by the inherently statistical nature of large language models (LLMs), by integrating structured knowledge and symbolic reasoning, while also supporting evidence derived from both text and images.

Research challenges / Novelty / Innovation
• Integration of structured knowledge
• Domain-adaptable data-level training recipes
• Agentic compositional reasoning
• Development of a use case: a medical assistant with access to knowledge about retinal diseases and multiple test images per patient

Research results:
• Original algorithms and methods
• Demonstrator

Innovate:
• Original neuro-symbolic solutions