📝 SummarXiv

Abstract

Recurrent-depth transformers scale test-time compute by iterating latent computations before emitting tokens. We study the geometry of these iterates and argue for a simple, \emph{two-scale} operational picture: (i) within a looped block, updates act as \emph{small-scale refinements}; (ii) across consecutive blocks, states undergo a \emph{larger-scale drift}. Across checkpoints, our measurements show that loop steps become \emph{smaller} and increasingly \emph{orthogonal} to one another, indicating better local modeling of fine structure rather than merely pushing in a single direction. These dynamics motivate an early-exit mechanism based on the model's second-order difference in step-size, which we show is superior in terms of performance, stability and time-efficiency, when compared to the KL-divergence exit strategy of Geiping et al. and its naive first-order counterpart.