A new paper studies depth-recurrent transformers, a class of models that repeatedly applies a weight-tied core rather than using a fixed stack of separate layers. The authors focus on whether individual tokens converge as the recurrence proceeds.
The result is useful for model builders exploring adaptive compute, recurrent inference, and architectures that spend more work only where it helps. Understanding convergence behavior can make these systems easier to reason about and potentially cheaper to run.
The paper adds to a broader research push around making transformer computation more flexible without giving up the modeling gains that come from deep iterative processing.