IAAIAgreement, Not Normalisation
Agreement rather than normalisation keeps depth reusable in looped language models
Asks what keeps extra loop steps usable beyond the trained budget. Removing Ouro's loop-boundary normalisation breaks the released models without norm blow-up, yet small models trained without it extrapolate better, suggesting that agreement between the training and inference loop is what matters.
Inside the method
[ Block ] × R with or without loop-boundary normalisation
Simplified conceptual schematic. Consult the paper for the complete architecture.
- Recurrence family
- Full-stack recurrence
- Depth control
- Extrapolation beyond the trained loop budget
- KV / state strategy
- See paper
Reading note
The damage concentrates in the readout and last block layer. At matched overshoot of the trained budget, the advantage of looped over stacked depth vanishes on one synthetic task and shrinks on the other.
Sources checked 2026-10-28. This catalog does not imply independent reproduction.
Related works
Cite this work
@misc{loopagreement2026,
title = {Agreement rather than normalisation keeps depth reusable in looped language models},
author = {Ivan Viakhirev and Kirill Borodin and Maxim Abramov and Grach Mkrtchian},
year = {2026},
eprint = {2610.31943},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.31943}
}