IAAITwo-Scale Latent Dynamics
Two-Scale Latent Dynamics for Recurrent-Depth Transformers
Describes recurrent-depth iterates as small refinements within a looped block and larger drift across blocks; loop steps grow smaller and more mutually orthogonal over training. This motivates an early exit based on the second-order difference of step size.
Inside the method
[ Looped block ] × r → exit when second-order step-size difference is small
Simplified conceptual schematic. Consult the paper for the complete architecture.
- Recurrence family
- Adaptive depth
- Depth control
- Early exit from latent geometry
- KV / state strategy
- See paper
Reading note
The second-order exit beats Huginn's KL-divergence exit and a first-order variant in performance, stability and time efficiency.
Sources checked 2026-10-28. This catalog does not imply independent reproduction.
Related works
Cite this work
@misc{twoscalelatentdynamics2025,
title = {Two-Scale Latent Dynamics for Recurrent-Depth Transformers},
author = {Francesco Pappone and Donato Crisostomi and Emanuele Rodolà},
year = {2025},
eprint = {2509.23314},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2509.23314}
}