"Theater of Mind" for LLMs: A Cognitive Architecture Based on Global Workspace Theory

Researcher Wenlong Shang has introduced a novel cognitive architecture for Large Language Models (LLMs) called Global Workspace Agents (GWA), detailed in the paper "Theater of Mind' for LLMs: A Cognitive Architecture Based on Global Workspace Theory." The work addresses a fundamental limitation of current LLMs, which operate as Bounded-Input Bounded-Output (BIBO) systems—remaining passive until prompted and lacking intrinsic temporal continuity. Shang argues this reactive paradigm is a bottleneck for true AI autonomy. Current multi-agent frameworks often fail during extended tasks due to static memory and passive communication, leading to cognitive stagnation.

GWA is inspired by Global Workspace Theory from cognitive science, which posits a central 'theater' of consciousness where information is broadcast. The architecture transforms multi-agent coordination into an active, event-driven dynamical system. It features a central broadcast hub that manages a heterogeneous swarm of functionally specialized agents, maintaining a continuous cognitive cycle. A key innovation is an entropy-based intrinsic drive mechanism that mathematically quantifies the semantic diversity of the system's outputs. This allows it to dynamically regulate the LLM's generation temperature to autonomously break reasoning deadlocks when thought becomes too repetitive or stuck.

Furthermore, the architecture employs a dual-layer memory bifurcation strategy to separate and manage different types of information, ensuring long-term cognitive continuity across tasks. Unlike frameworks that simply chain prompts, GWA provides a structured, reproducible engineering blueprint for building LLM-based systems capable of sustained, self-directed agency without constant human intervention.