Study shows verbalizable representations create a global workspace in language models

A new research paper examines how language models organize internal information. The authors propose that verbalizable representations serve as a global workspace. This workspace enables different parts of the model to share knowledge. The study analyzes transformer circuits to identify such representations. Findings suggest that verbalizable vectors can be accessed across layers. The work contributes to understanding interpretability of large models. The authors discuss implications for model design and debugging. The paper is available on the Transformer Circuits website.