Verbalizable Representations Operate as a Global Workspace in Language Models

A new research paper examines the internal dynamics of language models. The authors argue that verbalizable representations act as a shared workspace. This workspace allows different model components to exchange information. The concept draws parallels to cognitive theories of global workspaces. Experiments demonstrate that models can coordinate via these representations. Findings suggest a pathway toward more interpretable AI behavior. The work contributes to ongoing debates on model transparency. Future studies may explore how to harness this mechanism for controlled generation.