A global workspace in language models

Anthropic researchers have identified a specific internal structure within large language models that functions similarly to a global workspace, allowing information to be shared across the system. By mapping how these models integrate data, the team demonstrated that they can isolate and influence specific concepts, such as identifying a lie or recalling a fact. This finding offers a more precise method for interpreting model behavior, potentially improving safety and control by revealing how neural networks process and prioritize information during decision-making.