Anthropic Research finds verbalizable representations form a global workspace in language models

Anthropic researchers have published a study demonstrating that verbalizable representations within large language models function as a global workspace. This work investigates how specific internal model states facilitate the sharing of information across different cognitive processes.

The research identifies that these verbalizable representations act as a central hub, allowing disparate parts of the model to access and utilize shared context effectively. This mechanism supports complex reasoning tasks by enabling a unified representation of the current state.