WorkspaceBench: Evaluating Interpretability Methods for the Global Workspace

TL;DRWe introduce WorkspaceBench, a set of evaluations for how well an activation-to-text tool can read the contents of the “global workspace” of a model, i.e.