Reusing a single Transformer block across multiple depth steps is a known way to increase computation without duplicating parameters. But as the authors note, successive passes have no dedicated state through which positions can exchange compact summaries of their causal prefix.
The paper proposes "causal latent workspaces" as a solution—an explicit state that persists across loop iterations, allowing positions to store and share summaries of what they have seen. This is intended to give looped transformers a form of memory that standard weight-sharing lacks.
The available abstract does not include experimental results or implementation details, so the claims are limited to the conceptual proposal. The work is presented as a new direction for designing looped architectures.