Saved in:
Bibliographic Details
Main Authors: Pappone, Francesco, Crisostomi, Donato, Rodolà, Emanuele
Format: Preprint
Published: 2025
Subjects:
Online Access:https://arxiv.org/abs/2509.23314
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866918200001691648
author Pappone, Francesco
Crisostomi, Donato
Rodolà, Emanuele
author_facet Pappone, Francesco
Crisostomi, Donato
Rodolà, Emanuele
contents 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, two-scale operational picture: (i) within a looped block, updates act as small-scale refinements; (ii) across consecutive blocks, states undergo a larger-scale drift. Across training, our measurements show that loop steps become smaller and increasingly 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.
format Preprint
id arxiv_https___arxiv_org_abs_2509_23314
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Two-Scale Latent Dynamics for Recurrent-Depth Transformers
Pappone, Francesco
Crisostomi, Donato
Rodolà, Emanuele
Machine Learning
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, two-scale operational picture: (i) within a looped block, updates act as small-scale refinements; (ii) across consecutive blocks, states undergo a larger-scale drift. Across training, our measurements show that loop steps become smaller and increasingly 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.
title Two-Scale Latent Dynamics for Recurrent-Depth Transformers
topic Machine Learning
url https://arxiv.org/abs/2509.23314