Q-CHARM: Quantum-Compositionality in Human-like Abstraction, Reasoning and Machine Learning

The recent surge in artificial intelligence research and applications has reignited interest in
developing systems capable of human-like learning and reasoning while remaining transparent,
interpretable and trustworthy. The overarching objectives of compositional AI research are twofold:
(1) to address the research debt in the field of AI by providing a rigorous unified category-theoretic
foundation encompassing the subfield, and (2) to craft hybrid models rooted in this foundation that
are inherently compositional. These models leverage the strengths of both symbolic and subsymbolic
representations to equip artificial agents with the ability to learn and reason about the world in a way
more reminiscent of human cognition.
The Q-CHARM project focuses on the role of compositional model design in enhancing compositional
behaviour, distinguishing between the structural composition of a model’s architecture, defined pretraining,
and the emergent compositional properties of its learned representations. By investigating
how structured biases shape learning efficiency, interpretability, and creative generalisation, QCHARM
aims to provide a rigorous foundation for compositionally-interpretable AI systems. Through
theoretical analysis, model development, and real-world deployment in a creative system, the project
seeks to demonstrate that explicitly structured models foster more interpretable, generalisable, and
creatively expressive representations.