DeepThought: An Architecture for Autonomous Self-motivated Systems

Arlindo L. Oliveira, Tiago Domingos, Mário A. T. Figueiredo, Pedro U. Lima

Abstract

Current Large Language Models (LLMs) are often conflated with autonomous agents; however, their stateless, autoregressive nature lacks the fundamental requirements for genuine agency and self-motivation. We propose DeepThought, a biologically inspired architecture that moves beyond monolithic models. By integrating LLMs as specialized linguistic modules within a broader framework of vision, memory, and attention controllers, DeepThought provides a blueprint for systems capable of meta-cognition and persistent internal drive.

Introduction

The AI community continues to debate whether scaling LLMs will lead to emergent consciousness or true agency. This paper argues that scaling alone is insufficient. LLMs are essentially passive predictors with fixed weights and finite context windows that reset per session. True agency requires a transition to autonomous agents that can set independent goals, learn continuously from interactions, and introspect on their own cognitive processes. DeepThought addresses this gap by proposing a modular "System 1 / System 2" approach rooted in cognitive neuroscience.

The DeepThought Architecture

The DeepThought framework decomposes intelligence into specialized, interacting modules rather than a single neural network.

The DeepThought architecture integrates multimodal inputs, an embeddings store, and a high-level supervisor module for autonomous control.

Key components of the architecture include:

  • Supervisor Module: Functioning as a "System 2" controller, it manages short-term working memory and directs the system's focus.
  • Embeddings Store: A multimodal hub where sensory inputs (text, images, sound) are translated into a shared mathematical representation, facilitating cross-modal reasoning.
  • Complementary Learning Systems (CLS): Mimicking the human brain, the system utilizes a "fast-learning" component for immediate episodic experiences and a "slow-learning" component for long-term structural knowledge.

Cognitive Foundations for Agency

DeepThought leverages established theories from neuroscience to enable meta-cognition and self-directed behavior.

Global Neuronal Workspace (GNW)

Information from various sensory and cognitive modules is "broadcast" to a central workspace. This mechanism ensures that only the most relevant information is selected for high-level processing, effectively creating a bottleneck that mimics conscious attention.

The Global Neuronal Workspace model illustrates how specialized systems (memory, perception, value) interact through a central broadcast hub.

Attention Schema Theory (AST)

By maintaining a simplified internal model of its own attention, the agent "knows" what it is focusing on. This self-referential loop provides the foundation for meta-cognition, allowing the system to monitor and adjust its own processing strategies.

The Libido Sciendi

To drive autonomous behavior, the architecture incorporates an inherent "passion for knowledge." This is operationalized through principles like Friston’s Free Energy Principle (minimizing surprise) and Popper’s Falsificationism, encouraging the agent to actively test hypotheses about its environment.

Discussion and Future Work

DeepThought represents a theoretical roadmap rather than an empirical study. While it provides a blueprint for agency, several challenges remain. The "binding problem"—how modules communicate without information loss—and the precise training mechanisms for the "slow-learning" neocortical components require further research. Future work will focus on implementing this architecture in simulated environments to validate the emergence of goal-splitting and meta-cognitive properties.

Summary updated

BibTeX
@inproceedings{oliveira2023deepthought,
  author = {Arlindo L. Oliveira and Tiago Domingos and Mário A. T. Figueiredo and Pedro U. Lima},
  title = {{DeepThought: An Architecture for Autonomous Self-motivated Systems}},
  booktitle = {IMOL @ NeurIPS},
  year = {2023},
  url = {https://openreview.net/forum?id=RKGLAGayrt},
}
Figure