Active Inference Framework for Phenotyping AI Agency
A new study posted on arXiv (2604.23278) presents a way to define agency in AI using active inference. The researchers suggest a more straightforward view of agency, focusing on three main components: intentionality, which refers to actions influenced by beliefs and desires; rationality, which means actions based on a coherent understanding of the world; and explainability, where actions can be linked back to internal states. They frame these ideas within a partially observable Markov decision process in a variational context, where beliefs and preferences help form action sequences. Their experiments with a T-maze show that empowerment, indicated by the connection between actions and observations, distinguishes between zero agency and intentional agency. This study addresses a gap in understanding agency in the growing field of AI.
Key facts
- Paper published on arXiv with ID 2604.23278
- Proposes active inference for phenotyping agency in AI
- Three criteria: intentionality, rationality, explainability
- Uses partially observable Markov decision process
- Variational framework with posterior beliefs and prior preferences
- Empowerment metric based on channel capacity
- T-maze paradigm used for demonstration
- Distinguishes zero- from intentional agency
Entities
Institutions
- arXiv