STARS Framework Stabilizes Recurrent Dynamics in Looped Language Models

A new training framework called STARS (STAbility-driven Recurrent Scaling) addresses the instability of test-time scaling in Looped Language Models (LoopLMs). LoopLMs use depth recurrence for latent reasoning but often see performance peak then collapse with further iteration. STARS constrains latent states to approach asymptotically stable fixed points via Jacobian Spectral Radius Regularization with random loop sampling. Experiments on arithmetic tasks show STARS achieves stable and effective scaling, enabling longer reasoning chains without degradation. The work conceptualizes reasoning as uncertainty reduction and targets convergence to stable fixed points.