Incorporating the human perspective of attention into AI models

Develop AI modeling frameworks that explicitly integrate the human perspective of attention—particularly intentional control and joint attention—into Transformer-based systems, so that attention functions as an agency-related mechanism rather than solely computing pairwise relationships within input sequences.

Background

The authors argue that, unlike human attention which serves as an explicit component of agency and supports joint attention in social cooperation, Transformer-based models primarily use attention to compute relationships among sequence elements and do not encode intentional or joint-attentional mechanisms.

They note that recent multi-agent algorithms based on Transformers do not explicitly incorporate joint attention, motivating the need for approaches that bring human agency-related attention constructs into AI systems.