acceptodds
Under review as a conference paper at ICLR 2027

Shared Computation, Different Readouts: Depth Compatibility in Looped Transformers

Abstract

Multi-depth training can improve a looped transformer's accuracy under a shared output head, yet this advantage can reverse when both models receive a small adaptive interface. We study depth compatibility through the excess prediction risk of sharing one readout across depths instead of fitting equally expressive heads separately. In controlled in-context linear regression, we compare fixed-depth and multi-depth training with matched initializations, examples, optimizer updates, and cumulative recurrent computation. Multi-depth training lowers shared-head error but raises the error of a separately fitted head at the reference depth. After we freeze the cores, two shared adaptation directions remove over 90% of the all-depth sharing cost in each of six models. Four directions reverse the training comparison in all three initialization pairs on two separate 4,096-episode evaluations. Depth-specific gain and offset recover substantially less accuracy, even when we jointly optimize the common prediction direction. We characterize capacity-dependent risk under different, possibly singular feature geometries across depths. The task objective separates separately decodable risk from compatibility cost. An exact information-preserving recurrence shows that a large sharing cost can require only two adaptation directions. The comparison reverses with far less flexibility than fully independent heads provide. These results distinguish the prediction accuracy available in recurrent states from their compatibility with a shared output interface.

Primary area:
interpretability and explainable AI
Keywords:
looped transformers, representation interpretability

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_F81rOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper study depth compatibility in looped transformers. It separate prediction accuracy and shared readout cost. Authors propose risk decomposition and low-rank adaptive heads. Experiments on synthetic linear regression show multi-depth training better with shared head. But fixed-depth model become better with few adaptive directions.

Critical strengths

Good theoretical formulation. Paper clearly separate independently decodable prediction risk and shared-head compatibility cost. Risk decomposition is clear and useful. It explain why multi-depth training improve shared-head accuracy.

Experiment design is careful. Only four adaptive directions reverse comparison in all three pairs. This is interesting finding.

Critical weaknesses
  1. Limited experimental generalization. Main experiment only use small synthetic in-context linear regression, with four-dimensional input and width-32 transformer. This task are much simpler than real language modeling or reasoning. Although paper include Huginn model audit, it does not directly validate proposed adaptive readout mechanism. So not clear whether same conclusion hold for larger models and complex tasks.

  2. Limited training seeds and practical evaluation. Main comparison only have three paired initialization seeds. The confidence interval mainly resample test episodes, not independent trained models. Also adaptive heads are fitted using labeled examples from all 16 evaluated depths. This require extra supervision and depth-specific coefficients. Paper not fully show whether method can work efficiently for unseen depths or practical adaptive computation. Therefore practical significance is still uncertain.

  3. it is better to have a framework figure.

Questions for authors

Can authors test on larger transformer or real language tasks? Does same reversal still happen?

How sensitive is result to model width, training depth and number of seeds?

Can adaptive heads generalize to unseen depths without new labels?

can the authors provides with one first image in the paper as a framework pipeline figure?

Reviewer_85oAOpenReview ↗
Rating
4
Confidence
2
Informativeness
3
Presentation
1
Soundness
3
Read review
Paper summary
  • The authors argue that for looped transformers, task accuracy at different numbers of loops confounds two properties of the model: genuine improvement in prediction accuracy and the same information being easier to read at any loop count.
  • Their analysis hinges on studying the two in isolation. To this end, the authors decompose the risk into three terms, which allow independent analysis of the two conflated properties: "the risk achievable with independent heads, the cost of sharing a head, and the current head’s gap from the best common head".
  • This breakdown is operationalised by fixing the core looped transformer model and varying readout head variants. Specifically, the authors compare a single readout for all loop counts, separate readout heads, and an intermediate family of readout models in which the heads have r additional learnt read adaptation directions. While these directions are shared across depths, each loop count modulates the adaptation directions via per-depth coefficients.
  • The authors introduce their setup with mathematical rigour and demonstrate the decomposition on frozen looped transformers trained in an in-context linear regression setting.
Critical strengths
  1. Rigour and motivation

Overall, the paper is very rigorous and well-motivated. Authors are explicit in their assumptions and very careful in the claims they make, which, up to my understanding of the paper, I find correct. Decomposition of risk into three terms provides convenient objects to study. Additionally, in context linear regression is a fitting testbed to measure the decomposition, ensuring the target is exact and identical at every depth. Analysis is good and authors ran random controls to determine if the learnt directions provide any uplift over additional capacity.

Critical weaknesses
  1. Clarity of the manuscript

The manuscript is difficult to read, and the authors frequently use terms they haven’t defined yet in the paper (eg interface, adaptation direction). They often use multiple names for the same thing (head, readout, interface, probe), to the point that readers can easily get confused. Unlike the introduction, the abstract is very dense and relies on terms defined later in the main body, making it unhelpful for understanding the core of the paper. Minor issues: Table 1 is missing robustness statistics; Figure 1a has unlabeled horizontal lines.

Questions for authors
  1. [clarification] Choice of readout-relaxation family

The authors ask the core research question, “Does a training advantage survive when both frozen cores receive the same small increase in output-interface flexibility?” Why have the authors chosen this specific way to relax the shared-readout constraint at evaluation time?

  1. [clarification] Figure 1

Figure 1a is missing a legend for the dashed and dotted lines. The adaptation directions in Figures 1a and 1b are only shown for r in {0, 1, 2, 4}. I would appreciate it if the authors extended the support to the full range over which they varied r or included an explanation for why they do not show results for the larger loop counts in this figure.

Reviewer_Tor5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies whether performance differences between fixed-depth and multi-depth training of looped transformers reflect separately decodable prediction quality or compatibility with a shared output head. It decomposes prediction risk into independent-head risk, a sharing cost, and the native head’s excess risk, and studies depth-dependent affine heads of the form w_t = w + Ua_t.

In matched in-context linear regression experiments, multi-depth training performs better with a shared head, whereas fixed-depth training achieves lower error after both frozen cores receive a small adaptive interface. Two learned directions remove over 90% of the estimated sharing cost in all six matched models, and four directions reverse the comparison in all three initialization pairs on two evaluation sets. Random-direction and scalar-calibration controls support the importance of adapting the prediction directions.

Critical strengths

The frozen-core, equal-capacity comparison provides a useful diagnostic beyond independent probing. The paper demonstrates that the ordering of two training regimes changes under the same small relaxation of the readout constraint. Because recurrent states remain fixed, this directly supports the claim that the output interface materially affects the comparison. Random directions and both frozen and jointly optimized scalar-calibration controls make the result more informative than simply showing that separately fitted heads improve accuracy.

The analysis carefully distinguishes mathematical identities from empirical estimates. The treatment accommodates depth-dependent and singular feature geometries, while the experimental discussion acknowledges regularization, finite-sample effects, and nonconvex fitting. Independent confirmation episodes and per-pair results strengthen the reported reversal, and the uncertainty intervals are correctly described as conditional on the trained cores.

Critical weaknesses

The principal empirical finding is established in a narrow setting. The matched reversal uses three initialization pairs of width-32 models on four-dimensional, scalar-output linear regression. The additional configurations do not provide a corresponding matched comparison at larger scale, and the Huginn audit does not test the frozen-state capacity intervention. Since the main empirical contribution is that very few adaptation directions suffice to reverse a training comparison, it remains unclear how strongly this finding depends on the task, architecture, training duration, and evaluation-depth distribution.

The analysis explains the observed ranking more convincingly than it explains the training tradeoff. The objective decomposition is an accounting identity, and the matched training comparison changes supervision depth and gradient structure together. It therefore does not establish whether the fixed-depth core’s lower independent-head risk results from avoiding a compatibility constraint, concentrating supervision at depth four, or differences in optimization progress. The manuscript acknowledges the absence of causal mediation identification, but this leaves the training-mechanism insight limited. A comparison under the same multi-depth schedule with different training-time readout constraints would substantially strengthen that interpretation.

Questions for authors

Robustness of the central finding: Does the equal-capacity reversal persist when results are reported separately over depths 1–7 and 8–16, and across longer training budgets or additional paired seeds? These checks would help determine whether the conclusion is sensitive to the chosen evaluation distribution or optimization stage. Role of the training-time sharing constraint: Under the same balanced depth schedule and comparable training resources, what happens when the core is trained with a small adaptive interface or independent heads instead of one shared head? Does relaxing that constraint reduce the independent-head risk gap relative to fixed-depth training?

Scope of low-dimensional adaptation: Can the authors provide a matched fixed-versus-balanced comparison for the existing d=8 setting, reporting both capacity curves and per-seed reversals? This would directly test whether the principal result extends beyond the main four-dimensional configuration.

Reviewer_An5YOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates whether the benefits of multi-depth training in looped transformers arise from improved predictive information in recurrent states or better compatibility with a shared output head. The authors formalize depth compatibility as the excess prediction risk of using a shared affine readout instead of independently optimized depth-specific heads. They derive a risk decomposition and characterize how low-dimensional adaptive readouts can reduce this compatibility cost, including under depth-dependent and singular feature geometries.

Critical strengths
  1. The paper identifies an important confounding factor in evaluating looped transformers: improved performance under a shared output head does not necessarily imply better recurrent computation. The proposed risk decomposition provides a principled framework for separating independently decodable prediction accuracy from cross-depth compatibility costs. This offers a useful perspective on interpreting multi-depth training and recurrent-depth scaling.

  2. The authors carefully match initialization, training data, optimization steps, and cumulative recurrent computation. The finding that only a few learned adaptation directions can reverse the performance comparison between fixed-depth and multi-depth training is particularly interesting. The theoretical characterization of low-dimensional compatibility costs with calibration controls and independent test evaluations.

Critical weaknesses
  1. The main experiments are conducted on small synthetic in-context linear regression tasks with a width-32 transformer and only three paired initializations. Although this controlled setting is appropriate for isolating readout compatibility, it remains unclear whether the observed low-dimensional structure and reversal of training advantages generalize to large-scale language modeling. In particular, practical looped architectures may include additional non-shared layers between the recurrent computation and the LM head, potentially changing the nature of the readout compatibility problem.

  2. The proposed readouts are fitted using additional labeled data from all evaluated recurrence depths, while the recurrent backbone remains frozen. Therefore, the experiments primarily demonstrate that useful predictions can be recovered retrospectively, rather than establishing a practical mechanism for improving inference-time depth generalization. It would be valuable to investigate whether the learned adaptation directions generalize to unseen recurrence depths and whether similar improvements can be achieved with lightweight interfaces under realistic deployment constraints.

Questions for authors
  1. The main findings are established on small synthetic linear regression tasks. It remains unclear whether the same readout compatibility issue persists in realistic looped language models, particularly architectures with additional nonlinear layers (e.g., a coda) between the recurrent block and the LM head. Can the authors clarify the expected applicability of their findings to such architectures? Is there evidence that similarly low-dimensional readout adaptations can recover prediction accuracy in pretrained looped language models?

  2. The adaptive readouts are fitted using labeled examples from all 16 evaluated depths. Can the authors investigate whether adaptation directions learned from a subset of depths generalize to unseen depths, especially beyond the training range? This distinction seems important for understanding whether the proposed low-dimensional structure reflects a general property of recurrent representations or primarily facilitates fitting the evaluated depths.

  3. The main comparison uses only three paired initialization seeds, and the reported confidence intervals condition on the trained models. Given the relatively small performance difference after adaptation, can the authors provide additional evidence of robustness across independently trained models? In particular, how consistently does the performance reversal occur with different initializations and training configurations?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.