Looped Transformers under the Jacobian Lens: Does the Global Workspace Survive Recurrence?
Extends the Jacobian lens to iterated/recurrent transformer architectures via a virtual-unrolling adapter, testing whether a global workspace emerges when depth is implemented through recurrence rather than distinct layers Demonstrates that a workspace does form in the iterated portion of both looped (Ouro-2.6B) and depth-recurrent (Huginn-0125) architectures, but recurrence fundamentally alters how workspace content can be accessed and transported Ouro-2.6B (48 layers looped 4 times) reconstruc
Analysis
TL;DR
- Extends the Jacobian lens to iterated/recurrent transformer architectures via a virtual-unrolling adapter, testing whether a global workspace emerges when depth is implemented through recurrence rather than distinct layers
- Demonstrates that a workspace does form in the iterated portion of both looped (Ouro-2.6B) and depth-recurrent (Huginn-0125) architectures, but recurrence fundamentally alters how workspace content can be accessed and transported
- Ouro-2.6B (48 layers looped 4 times) reconstructs workspace content in every loop, but linear transport cannot carry content across loop boundaries, requiring writes and ablations to span every remaining loop
- Huginn-0125 (4-layer core recurred 16 times) carries content forward across all recurrences, while reads, writes, and ablations operate only within a sliding window of roughly two recurrences
- Verbalization of newly injected content tracks explicit per-iteration supervision, whereas steering of existing content does not depend on such supervision
Why It Matters
This work directly addresses a critical architectural question for efficient AI systems: whether the functional properties of deep transformers (specifically the global workspace phenomenon) survive when depth is compressed through recurrence. For practitioners building parameter-efficient models or exploring recurrent alternatives to deep feedforward stacks, understanding workspace accessibility constraints is essential for designing effective intervention and interpretability pipelines.
Technical Details
- Virtual-unrolling adapter: The authors extend the Jacobian lens methodology to iterated architectures by introducing a virtual-unrolling adapter that enables analysis of recurrent/looped models without full unrolling, making causal analysis computationally tractable.
- Three models compared: Ouro-2.6B (48 layers looped 4 times with deep supervision), Huginn-0125 (4-layer core recurred 16 times, trained for latent reasoning), and Qwen3.6-27B (64 untied layers as the standard feedforward baseline).
- Full workspace suite applied: Lens fitting, readout analysis, and eleven causal experiment families were used to characterize workspace formation and accessibility across all three architectures.
- Key finding on transport: In Ouro, linear transport fails across loop boundaries—writes and ablations must span every remaining loop to be effective. In Huginn, content propagates across all 16 recurrences, but causal operations (reads, writes, ablations) are confined to a sliding window of approximately two recurrences.
- Supervision dependency: Whether newly injected content can be verbalized depends on explicit per-iteration supervision, but the ability to steer existing workspace content is independent of such supervision.
Industry Insight
- Model compression through recurrence is viable for preserving workspace functionality, but interpretability and intervention tools must account for recurrence-specific access constraints—standard linear probing techniques may fail across loop boundaries in looped architectures.
- The sliding-window causal effect in Huginn suggests that for deeply recurrent models, intervention strategies should target recent recurrences rather than attempting to reach earlier states, which has direct implications for mechanistic interpretability pipelines on efficient recurrent models.
- Per-iteration supervision is a key lever for controlling verbalization of injected content but not for steering existing representations, suggesting that training objectives should be carefully designed depending on whether the goal is content injection or content manipulation in recurrent systems.
Disclaimer: The above content is generated by AI and is for reference only.