Do looped transformers have a global workspace?

Yes. New research using the Jacobian lens on Ouro-2.6B and Huginn-0125 shows that looped transformers develop a mid-depth band of verbalisable, causally potent representations — a functional analogue of a global workspace — despite reusing weights across depth.

How does recurrence change workspace access compared to standard transformers?

Access patterns depend on the loop design. In Ouro-2.6B (48 layers looped 4x), workspace content is reconstructed in every loop and cannot cross loop boundaries via linear transport. In Huginn-0125 (4-layer core recurred 16x), content persists across all recurrences but interventions only affect a sliding window of roughly 2 recurrences.

What is the Jacobian lens and how was it used in this study?

The Jacobian lens is a mechanistic interpretability technique that tracks how perturbations at one layer propagate through a neural network. The authors extended it to iterated architectures via a virtual-unrolling adapter, enabling them to apply lens fitting, readout probes, and 11 families of causal experiments on looped transformers.

Why does the global workspace matter for AI safety?

The workspace represents the model's causally potent, verbalisable internal state. Knowing how to read, write, and ablate these representations is essential for activation steering and representation engineering — techniques used to control model behaviour and ensure safety. This paper shows these techniques transfer to recurrent architectures but need recalibration for each loop design.

Global Workspace Survives in Looped Transformers: 11 Causal Tests

Recent work found that standard feedforward transformers develop a mid-depth band of representations that are simultaneously verbalisable and causally potent — a functional analogue of a global workspace, the cognitive theory of conscious access. But that result came from architectures with distinct, untied weights per layer. What happens when depth is implemented through recurrence — reusing the same weights across multiple passes?

A new preprint from Wenlong Wang and Fergal Reid puts this question to the test. Using the Jacobian lens — a mechanistic interpretability technique that maps how representations change across layers — they apply a full suite of 11 causal experiment families to two looped architectures: Google DeepMind's Ouro-2.6B and Apollo Research's Huginn-0125, with Qwen3.6-27B as the standard baseline. The answer: a workspace does form under recurrence, but the rules of engagement change.

What Problem Does This Solve?

The global workspace hypothesis, borrowed from cognitive neuroscience, posits that conscious access requires a centralised, broadly accessible pool of information that can flexibly steer behaviour. In transformers, a mid-depth band of representations fits this description: you can read out their content, intervene on them causally, and those interventions predictably alter outputs.

But looped transformers — where the same weights are applied repeatedly — challenge this picture. If the workspace depends on distinct layers with specialised functions, reusing weights across depth might destroy it. Alternatively, recurrence might produce a workspace with different access properties: harder to write to, harder to read from, or confined to sliding windows. The question matters for two reasons. First, it tests whether the workspace is an architectural accident or a functional necessity of deep computation. Second, looped architectures are gaining traction for their parameter efficiency and latent reasoning capabilities — if their internal representations are less steerable, that has safety implications.

What's the Method?

The authors extend the Jacobian lens — which tracks how perturbations at one layer propagate through the network — to iterated architectures via a virtual-unrolling adapter. This lets them apply the same mechanistic interpretability toolkit that works on standard transformers: lens fitting (training linear probes to read out representations), readout (measuring what information is accessible), and 11 families of causal experiments (activating, ablating, and editing representations at specific depths).

They test two very different looped designs:

Ouro-2.6B — 48 layers looped 4 times (192 effective steps), with deep supervision at every layer boundary. Think of it as a standard transformer whose layers are partitioned into 4 repeating blocks.
Huginn-0125 — a 4-layer core recurred 16 times (64 effective steps), trained specifically for latent reasoning without explicit per-step supervision.

Qwen3.6-27B (64 untied layers) serves as the feedforward baseline, letting them isolate what recurrence changes versus what's inherent to depth.

mindmap
  root((Looped Transformers))
    Ouro-2.6B
      48 layers
      Looped 4x
      Deep supervision
      Workspace in every loop
      No linear transport across loop boundaries
    Huginn-0125
      4-layer core
      Recurred 16x
      Latent reasoning training
      Content persists across all recurrences
      Sliding window of ~2 for interventions
    Qwen3.6-27B Baseline
      64 untied layers
      Standard workspace access

What Are the Results?

Three main findings emerge from the 11 causal experiment families:

1. A workspace forms in both looped architectures. Despite reusing weights, both Ouro and Huginn develop mid-depth bands of verbalisable, causally potent representations. The workspace is not an artifact of untied layers — it's a functional consequence of sufficient depth, even when that depth is implemented cyclically.

2. Access patterns diverge sharply between the two designs. In Ouro, workspace content is reconstructed independently in every loop. Linear probes cannot carry information across loop boundaries — each 48-layer pass re-derives the workspace from scratch. This means that causal interventions (writes and ablations) must span every remaining loop to be effective; a single-loop intervention gets overwritten. In Huginn, by contrast, content flows across all 16 recurrences naturally. But interventions — reads, writes, and ablations — operate within a sliding window of only about 2 recurrences. You can steer the model, but only for the next ~8 effective steps before the effect dissipates.

3. Verbalisability tracks supervision; steerability does not. Whether newly injected content can be verbalised (read out by a probe) depends on whether the architecture received explicit per-iteration supervision during training — Ouro has it and shows better verbalisability; Huginn does not and shows worse. But whether existing workspace content can be causally steered (modified by interventions) is independent of supervision. This dissociation suggests that different mechanisms underlie reading and writing in the workspace.

What Are the Limitations?

The study compares only two looped architectures against one baseline, so the generality of the findings is uncertain. Ouro and Huginn differ in model scale, loop count, supervision regime, and training objective — any of these variables could drive the observed access differences. The Jacobian lens itself has assumptions (linear approximation of representation dynamics) that may break down at sharp nonlinearities. The causal experiments test intervention effects but do not establish that the workspace representations are necessary for the model's behaviour in a strong sense — only that they are sufficient to alter it. Finally, the study does not address whether workspace properties transfer across tasks or depend on the specific training distribution.

Why Should Someone Building Things Care?

If you're building with looped or recurrent transformer architectures — and the trend toward parameter-efficient, latent-reasoning models is accelerating — this paper tells you that the internal representations are steerable, but how you steer them depends critically on the loop structure. Interventions on Ouro require broad, multi-loop ablations. Interventions on Huginn are temporally local. This has direct implications for activation steering, representation engineering, and any safety technique that assumes uniform access to a model's internal state.

More broadly, the result that a workspace emerges under recurrence supports the view that it's a functional necessity of deep computation, not a quirk of untied weights. That's good news for mechanistic interpretability: the toolkit developed for standard transformers largely transfers to recurrent variants, even if the access patterns need recalibration.