August 21, 2025

Toward a reusable architecture for intelligent agents

Key Points

The proposed framework enhances generalisation across tasks in intelligent agents, making them more adaptable in diverse environments.
Key components include active inference and modularity, crucial for developing scalable generative models.
This work outlines a graph-like architecture that mirrors the connectivity and specialisation found in the human brain.
It suggests that implementing these core components could lead to the creation of more versatile intelligent systems.

Abstract

Active inference provides a powerful framework for understanding how agents perceive, learn, and act by maintaining an internal model of the world. However, most existing implementations rely on generative models that are designed for specific tasks or domains. These task specific models limit the ability of agents to generalise, adapt to new environments, or scale beyond narrow applications.Taking inspiration from the human brain, which appears to solve the problem of generalisation through a combination of modularity, hierarchy, and embodiment, we propose a list of core ingredients. We outline seven core components that we believe are essential for a scalable and reusable generative model, including compositional structure, temporal abstraction, attention and precision, meta-level learning, separation of self and world, and mechanisms for learning model structure itself.The discretisation of these core components hints at the necessity for a series of interconnected generative models organised on a graph-like structure with (some) shared variables, states and parameters, similar to the functional specialisation, integration, and connectivity of the brain. We propose that these components offer a blueprint for constructing intelligent systems that can generalise across tasks and domains.

Mark Helpful

Bookmark

Relay