New mechanisms for controllability and attention understanding in AI systems

Multimodal Large Language Models rely on attention mechanisms as a core component of Transformer architectures, yet attention remains largely implicit, limiting controllability, interpretability, and higher-level reasoning. Building on the EIC Pathfinder project ASTOUND and inspired by Theory of Mind and Attention Schema Theory, this project aims to make attention an explicit, modellable, and steerable process. The research will investigate a hierarchical two-level framework that leverages principles from flow- and diffusion-inspired generative modelling to predict and regulate attention dynamics in state-of-the-art pre-trained Transformers. By enabling deliberate control over attention allocation, the proposed approach seeks to enhance reasoning and planning capabilities while improving explainability, contributing to the development of more transparent, trustworthy, and cognitively grounded AI systems.