Recent work has identified a mid‑depth band of verbalisable, causally potent representations in a standard feed‑forward Transformer, serving as a functional analogue of a global workspace. It remains unknown whether the same workspace functionality emerges when depth is realized through recurrence rather than a stack of distinct layers. Looped and depth‑recurrent Transformers provide a direct test because they reuse the same weights across depth. We extend the Jacobian lens to iterated architectures via a virtual‑unrolling adapter and apply the full workspace suite—lens fitting, readout, and eleven causal experiment families—to three models.
The models are Ouro‑2.6B (48 layers looped 4 times, deeply supervised), Huginn‑0125 (a 4‑layer core recurred 16 times, trained for latent reasoning), and Qwen3.6‑27B (64 untied layers) as the standard baseline. We find that a workspace forms in the iterated part of each architecture, but recurrence changes how it can be accessed. Ouro reconstructs workspace content in every loop; linear transport cannot carry that content across loop boundaries, so writes and ablative interventions must span all remaining loops. Huginn carries content forward across all sixteen recurrences, while reads, writes, and ablative interventions operate only within a sliding window of roughly two recurrences.
Moreover, the ability to verbalise newly injected content tracks explicit per‑iteration supervision, whereas steering existing content does not.
Review