IAAIJacobian Lens on Looped Models
Looped Transformers under the Jacobian Lens: Does the Global Workspace Survive Recurrence?
Extends the Jacobian lens to iterated architectures to test whether a global-workspace-like band of verbalisable, causally potent representations survives recurrence, comparing Ouro-2.6B, Huginn-0125 and an untied Qwen baseline.
Inside the method
Virtual unrolling adapter → Jacobian lens over iterated layers
Simplified conceptual schematic. Consult the paper for the complete architecture.
- Recurrence family
- Full-stack recurrence
- Depth control
- Analysis of released models
- KV / state strategy
- See paper
Reading note
A workspace forms in the iterated part of both models, but access differs: Ouro rebuilds the content every loop, so writes must span all remaining loops, while Huginn carries it across all sixteen recurrences yet interventions act within about two recurrences.
Sources checked 2026-10-28. This catalog does not imply independent reproduction.
Related works
Cite this work
@misc{jacobianlenslooped2026,
title = {Looped Transformers under the Jacobian Lens: Does the Global Workspace Survive Recurrence?},
author = {Wenlong Wang and Fergal Reid},
year = {2026},
eprint = {2609.01924},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2609.01924}
}