acceptodds
Under review as a conference paper at ICLR 2027

Looped Transformers under the Jacobian Lens: Does the Global Workspace Survive Recurrence?

17 views

Abstract

A Jacobian lens identifies verbalisable, causally editable representations in standard transformers, but it is unclear whether the same workspace functions arise when depth reuses weights. We extend the lens to recurrent architectures by virtual unrolling and fit distance-controlled lens families for Ouro-2.6B and Huginn-0125, with Qwen3.6-27B as an untied baseline. Readouts, sparse concept inventories and eleven behavioural experiment families reveal two maintenance regimes. Ouro reconstructs content across supervised loop checkpoints: sustained writes outperform single-loop writes, while recovery from ablation depends on its width. Huginn's content remains readable across sixteen recurrences, but input-path interventions show that this persistence depends on continued re-injection; reads and effective edits follow a roughly two-recurrence horizon. Reporting existing and newly injected content also dissociates: Huginn excels at the former yet fails at the latter, which only Ouro achieves among the tested checkpoints. The results support functional workspace organisation under recurrence while distinguishing model behaviour from limitations of the readout instrument.

Primary area:
interpretability and explainable AI
Keywords:
Jacobian lens, looped transformers, depth recurrence, weight tying, global workspace, interpretability, latent reasoning, deep supervision, causal interventions, activation steering, representation transport, mechanistic interpretability

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_1xqmOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper investigates whether recurrent architectures (Ouro-2.6B and Huginn-0125, with Qwen3.6-27B as baseline) exhibit the same functional workspace organisation as standard transformers, using distance-controlled Jacobian lens analyses and eleven behavioural experiment families, and finds that functional workspace organisation is preserved under recurrence despite being partially masked by readout limitations.

Critical strengths

None

Critical weaknesses
  1. Lack of a clear research problem and effective contributions. After reading the Abstract, I did not obtain any substantive information about the proposed method or the core insight of the work. The author appears to merely describe a scenario involving a Jacobian Transformer and a set of specific experimental observations. Beyond that, it remains unclear what scientific problem the author intends to solve or what novel approach is being proposed. Furthermore, the wording throughout the Abstract is highly obscure and convoluted, making it very difficult for readers to follow the author's intended meaning.
  2. Weak motivation and unclear logical flow in the Introduction. The author raises the question "Does the same functional organisation arise when depth is implemented by re-applying the same weights?" However, it is unclear whether this question is sufficiently important, and the author does not review to what extent this question has already been addressed in existing work. No analysis or motivation is constructed around this question. The Introduction then abruptly moves on to describing two checkpoints the author examined, followed by an account of reproducing baseline methods. The overall narrative reads more like an undergraduate course lab report than a coherent research paper, leaving the reader confused about what the author actually aims to convey.
Questions for authors

This paper reads like an undergraduate course lab report and does not belong at ICLR. Course assignments should not be submitted to ICLR in the first place; doing so not only violates the spirit of the conference's submission policy but also severely wastes reviewers' time and resources.

2 comments

We have re-organised the paper to improve its clarity. Section, figure and table numbers are those of the revised paper.

W1. Research problem, approach and contributions.

Research problem. A looped transformer applies one block of layers several times in a forward pass. The output does not show what a later iteration does with the state it receives. In standard transformers, the Jacobian lens of Gurnee et al. (2026) reads the concepts held in an intermediate state as words, and edits along its directions change the model's answer. The introduction now states the research question explicitly, in italics: "what does the Jacobian lens show in a looped transformer, and where does the picture differ from a standard one?" (Section 1).

Approach. We keep the lens estimator of Gurnee et al. and add what a looped model requires:

  1. One lens fitted per iteration, with a test at Ouro's loop boundary that shows why a single lens fails (Section 3).
  2. Edits restricted to one loop or one window (Section 5.2).
  3. Interventions on the prompt that Huginn feeds in again at every recurrence (Section 4.2).
  4. Analyses of what a loop does with the state it is handed, compared with standard transformers (Section 6).

The experiments form two groups, reading (Section 4) and writing (Section 5), and each test is introduced with an example. In "Fact: The language spoken in the country where the Amazon River ends is", the intermediate concept (the bridge) is Brazil and the prompt never states it. The reading test asks whether the lens recovers Brazil. The writing test moves Brazil onto Mexico and succeeds when the answer becomes Spanish.

Contributions. The abstract and the introduction list three:

  1. A looped model needs one lens per iteration (Section 3). A single lens to the final layer fails in Ouro's earlier loops, and the failure lies in the lens. A mean Jacobian carries the bridge to its target on 90% of 81 questions inside a loop and on 19% across a loop boundary, although the state at the boundary still holds it (Table 1). Fitting one lens to the end of each iteration restores the readout.
  2. Ouro can be read and edited in every loop, Huginn only in a window (Sections 4 and 5). Each of Ouro's four loops has a band of layers in which unstated concepts can be read and an edit changes the answer. Huginn can be read only in a short window before each lens target and edited only in the last window, and its answer stays readable only while the prompt is fed in again at every recurrence.
  3. Ouro rebuilds its state in every loop (Section 6). This is our main finding. The first layers of each later loop write almost the negative of the state they are handed (cosine −0.99), and the loop then writes the content back, bringing up the candidate answers in the order of the first loop. An answer written into loop 2 is still the final answer in 18.7% of questions, against 56.1% and 74.4% for the same edit in Gemma 3 12B and Llama-3.2-3B, whether that answer is wrong or right (Figure 7). Among the models we tested, only Ouro rebuilds its state in this way.

On the wording, every term is defined where it first appears (Section 2 onwards).

W2. Motivation and earlier work.

Why the question matters. Looped transformers are an important research direction for language models: Ouro and Huginn are trained to continue their computation in the hidden state over several iterations before they answer. This reasoning cannot be inspected as written reasoning can. The Jacobian lens turns it into words, which is a first step towards overseeing what such models compute before they answer. The output does not show whether a later iteration carries the work of the one before it further or computes it again. Our third contribution answers this for Ouro: each loop rebuilds the state it is handed.

How far the question has been addressed (Section 1, paragraphs 2 and 3; Section 7). For standard transformers, Gurnee et al. (2026) read an intermediate state in words and test by edits whether the model uses what is read. To our knowledge, their lens has not been applied to looped models. Whether a looped model uses what can be read from its state, and what an iteration does with the state it is handed, remained untested.

On the remark under "Questions for authors".

The paper is not coursework. It is an interpretability study of two released looped language models, compared with three standard transformers. If the reviewer has further questions, we would be glad to discuss them.

Thank you for the response and clarification. I reread the revised abstract and found it clearer than the original. I appreciate the authors’ efforts. My current understanding is that the paper investigates what the Jacobian lens reveals about looped Transformers and how these findings differ from those for standard Transformers.

My first concern is that this question remains too broad. It reads more like an exploratory experiment than a well-defined research question. Almost any observations from these experiments could be presented as an answer. For a paper at a leading research conference, I would expect a more specific question and a clear explanation of why it matters. What makes this question worth studying, and how does it relate to prior work? Has it already been investigated?

The abstract states that the study compares two looped models with three standard Transformers. I do not find this range of models sufficient to support the conclusions being drawn. The authors report three findings, but in my view, these findings lack adequate theoretical justification and sufficiently comprehensive experimental evidence.

Finally, it is unclear how these observations translate into a substantive research contribution. Do they motivate a new method or help resolve a concrete problem? As presented, the paper reads largely as an experimental report, without a clear account of its scientific contribution. This is why I previously described it as resembling an undergraduate course lab report.

I will maintain my score, this is my final decision.

Reviewer_WapBOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper investigates whether recurrent, weight-tied transformers exhibit functional properties associated with a global workspace, adapting Jacobian-based interpretability methods to recurrent architectures. It studies Ouro and Huginn, using Jacobian-based readouts and behavioural interventions to examine the availability and reportability of internal content and its influence on behaviour. The reported observations include differences between the two models and a dissociation between reporting existing content and reporting newly injected content. However, the paper does not adequately define the functional properties that would constitute a workspace in recurrent models, or clearly map its experiments to justified criteria.

Critical strengths
  1. Potentially useful methodological adaptation. Adapting Jacobian-based interpretability methods and behavioural interventions to recurrent architectures could provide a useful way to investigate the function of internal representations in weight-tied models.

  2. Potentially informative empirical distinctions. The reported differences between Ouro and Huginn, and the dissociation between reporting existing content and newly injected content, are interesting observations that may help characterise different forms of content maintenance and access.

Critical weaknesses
  1. The central construct is insufficiently operationalised. The paper does not adequately explain what it means for a recurrent model to exhibit a global workspace, which properties are necessary for that interpretation, or how each experiment tests those properties. The theoretical connection to global workspace theory and to Gurnee et al. (2026) is not developed enough to support the main conclusion. Because this is the paper’s central question, the gap substantially limits the contribution.

  2. The exposition impedes evaluation of the argument. The manuscript does not consistently distinguish established concepts from newly introduced terminology and methods, nor explain them at the point where they are needed. This makes it difficult to reconstruct the reasoning and judge whether the conclusions follow from the evidence. In addition, differences in training, model scale, and iteration interfaces limit causal conclusions about architectural effects.

Questions for authors
  1. What specific properties do the authors regard as constitutive of a functional global workspace in a recurrent model, and which experiment or result provides evidence for each property? Please distinguish necessary criteria from suggestive indicators.

  2. How, precisely, does the operationalisation used here follow from the theoretical account in Gurnee et al. (2026) and the relevant global-workspace literature? Which aspects are inherited from that work, and which are new criteria introduced for recurrent architectures?

  3. Given the differences in training, model scale, and iteration interfaces between Ouro and Huginn, what conclusions can be supported about recurrence itself, as opposed to those models’ other differences?

4 comments

We thank the reviewer for the careful reading and for stating exactly what the argument lacked. Section, figure and table numbers below are those of the revised paper.

W1. The central construct is insufficiently operationalised.

We have re-organised the paper to improve its clarity. We answer the three parts of this point in turn.

What it means for a looped model to exhibit a workspace. Our question is whether the layers that the Jacobian lens of Gurnee et al. (2026) reads in a standard transformer also exist in a looped transformer, and how they differ there (Section 1). The lens gives us a way to understand looped transformers better, and it leads to our main finding, that Ouro rebuilds its state in every loop (Section 6).

Which properties, and how each experiment tests them. The experiments now form two groups, reading (Section 4) and writing (Section 5), and each test is introduced with an example prompt and its expected answer. In "Fact: The language spoken in the country where the Amazon River ends is", the intermediate concept (the bridge) is Brazil and the prompt never states it. The reading test asks whether the lens recovers Brazil. The writing test moves Brazil onto Mexico and succeeds when the answer changes from Portuguese to Spanish. Q1 below lists the properties and their tests.

The connection to global workspace theory and to Gurnee et al. Like Gurnee et al., we use "workspace" for functional properties only (Section 2.2). They write that they "do not claim that language models reproduce the full architecture global workspace theory ascribes to the brain", and neither do we; the word carries no claim about consciousness. Our connection to their work is the method. We apply their lens to looped transformers and find that the fit has to change, to one lens per iteration, before the states can be read clearly (Section 3).

Q1. Which properties, which evidence, and which are necessary.

We use the five properties by which Gurnee et al. define a workspace:

Property (Gurnee et al.)Our testResult
Internal reasoning: an unstated intermediate step drives the answerThe bridge of two-hop questions, read (Sections 3.1 and 4) and edited (Section 5.1, Figure 3)Read: Ouro puts the bridge first among control concepts on 90% of 81 questions inside a loop (Table 1). Edited: 47, 26 and 5 of 90 on Qwen, Ouro and Huginn
Flexible generalisation: one edited concept feeds many operationsThe argument of one-step questions moved onto another member of its category (Section 5.1, Figure 3)80, 73 and 25 of 192 trials, same order
Verbal report: the model names the conceptA swap of the model's own answer, and a word written in from outside (Appendix B.3, Table 4)Swap: 69%, 106 of 133 and 124 of 126. Written in: 0 of 101, 31 of 101 and 0 of 97
Directed modulation: an instruction changes what is readReadings under three instructions (Appendix B.2, Table 12)No consistent difference between the looped and the standard models
Selectivity: content is read only when the task needs itReadings with and without the task demand (Appendix B.2, Table 12)As above; readings only, with no edit

Gurnee et al. present the five as a cluster and do not mark any as necessary, and we follow them. Our claims that a model can be read and edited rest on the first two rows, which are also the tests that validate each lens (Section 3.2). The other three are supporting indicators, reported in the appendix. The dissociation the reviewer notes is the third row: all three models name a concept of their own after we swap it, and only Ouro names a word written in from outside. Ouro is also the only one of the three whose intermediate states are trained to be read by the output head, but it differs from the other two in scale and training recipe as well, so we do not attribute the result to that training (Section 8).

Q2. What is inherited from Gurnee et al. and what is new.

We take the following from Gurnee et al. without change: the lens estimator, the edits along lens directions, the 551 prompts with the concepts to recover, the five properties and their tests, and the name "workspace", which they take from global workspace theory (Baars, 1988). On our standard reference model the one-step edit succeeds on 42% of trials, close to the 40% they report for theirs.

The following are new for looped models:

  1. One lens per iteration, fitted on the unrolled computation, with every state read through the lens of its own iteration (Section 3.2, Appendix B.1).
  2. The test that shows why a single lens fails, on Ouro's loop boundary (Section 3.1, Table 1).
  3. Interventions on the prompt that Huginn feeds in again at every recurrence (Section 4.2).
  4. Edits restricted to one loop or one window (Section 5.2).
  5. The analyses of what a loop does with the state it is handed (Section 6).

W2, first part. Exposition.

The revised paper follows the three claims in order, and each term is defined where it is first used.

From earlier work we take the Jacobian lens, its target, reading a state, and edits along lens directions (Section 2.1), and the band and the workspace (Section 2.2). The terms we define ourselves are loop, recurrence, carried state and re-injected input, for the two looped models (Section 2.3); bridge, the unstated intermediate concept of a two-hop question (Section 3.1); argument tokens and final position (Section 5); rebuild (Section 6); margin and cancellation (Section 6.1); trap question (Section 6.2).

The terms the submission left undefined ("maintenance regime", "virtual unrolling", "interface") are no longer in the main text. The unrolling of the loop is described in Appendix B.1.

W2, second part, and Q3. What can be concluded about recurrence itself.

The differences in training, scale and iteration interface rule out conclusions about recurrence itself, and the revised paper draws none. Section 8 says: "The five models differ in scale and training recipe as well as in how they loop, so we describe each model and do not assign a difference between two models to a single design choice."

What the lens does show is how Ouro differs from standard transformers, and this is the main result of the revised paper (Section 6): at the start of every loop Ouro cancels the state it is handed, and then rebuilds it. We observe this in three ways.

  1. The state is cancelled and written back (Section 6.1, Figure 5). When loops 2 to 4 begin, the answer can already be read from the state handed over. The loop's first layers write almost the negative of that state, so the reading nearly disappears. The remaining layers then write the content back, along the same course as loop 1, which starts from nothing. Qwen, cut at mid-depth, shows no such cancellation: its later layers add to the state and leave it in place.
  2. Each loop goes through the candidate answers again (Section 6.2, Figure 6). On questions with a tempting wrong answer, a standard transformer brings up the tempting answer first and the correct one last, once. Ouro repeats this order in every loop, although loops 3 and 4 start from a state in which the correct answer already ranks first.
  3. An answer written into the state is replaced (Section 6.3, Figure 7). In Gemma 3 12B and Llama-3.2-3B, an answer written into the middle layers is usually the answer the model gives, because the later layers work from what they are handed. Written into Ouro's loop 2 it usually is not (18.7% of questions, against 56.1% and 74.4%), and it stays just as rarely when it is the right answer written into a question Ouro answers wrongly. Ouro's later loops therefore rebuild their own answer from the state they are handed, whichever answer the edit wrote. They still use that state: replacing the bridge in it breaks the answer.

Gemma 3 12B is close to Ouro's four loops in inference cost, Llama-3.2-3B is close in parameters, and both were pre-trained on a number of tokens of the same order as Ouro (12T and up to 9T, against 7.7T; Table 2, Appendix A). Among the models we tested, only Ouro rebuilds its state in this way.

For Huginn, one design feature is tested by intervention. Removing the re-injected input from recurrence 8 onwards moves the answer's median rank from 11 to 2,749 one recurrence later. Replacing it with the input of a prompt that differs in one word turns the output into the new prompt's answer in every case tested, even at recurrence 12 (Section 4.2). Huginn therefore computes its answer again from that input at each recurrence. Each model's pattern also holds when the same trained model is run with another number of iterations (2 to 6 loops for Ouro, 4 to 64 recurrences for Huginn; Appendices B.2 and B.3). These runs vary one model at inference and do not separate looping from training.

From Ouro against Huginn we draw no conclusion. Ouro was pre-trained on 7.7T tokens and Huginn on 0.8T. Huginn answers 25 of the 90 two-hop questions of Figure 3 unedited, against 43 for Ouro, and 10 of its 57 two-hop questions with a tempting answer, too few for the tests of Section 6. A difference between the two under the lens cannot be assigned to the way they loop, the training objective or the amount of pre-training. Whether Huginn would rebuild its state if pre-trained to Ouro's level is a question we cannot answer now.

Thank you for the detailed responses and for the substantial effort invested in revising the manuscript.

I have read the revised paper and appreciate the improvements in clarity and organization. My main concern in the original review was that Gurneeper did not sufficiently establish the connection between the functional global workspace studied by Gurnee et al. and the properties investigated in recurrent transformers. In particular, it was unclear how the experiments supported the claim that the same functional organization was being investigated.

The authors have addressed this concern primarily by substantially reframing the paper. Rather than attempting to establish whether the global workspace identified by Gurnee et al. persists under recurrence, the revised manuscript now investigates what the Jacobian lens reveals about recurrent transformers. I think this is a more appropriate framing for the experiments presented, and the revised research question is reasonable in its own right.

However, this change also substantially alters the scientific contribution of the paper. The first two contributions, concerning iteration-specific lenses and the different readability and editability patterns of Ouro and Huginn, are useful methodological and descriptive results. The third contribution, concerning the apparent reconstruction of representations across Ouro's loops, is more interesting and potentially significant for understanding the computation performed by recurrent models.

Nevertheless, I am not yet fully convinced by the interpretation of this third result. The experiments provide evidence for repeated cancellation and restoration of decodable representations, as well as for the tendency of later loops to override earlier interventions. However, these observations do not necessarily establish that the model recomputes the corresponding semantic content. In particular, geometric cancellation of a residual state does not imply that the information it contains has been erased, and the observed intervention effects do not uniquely identify the underlying computational mechanism. I appreciate that the authors acknowledge some of these limitations in the discussion.

More generally, I think the revised results point toward a promising research direction, but the central phenomenon would benefit from further investigation to establish its interpretation and generality. Since the focus of the paper has changed substantially, I also think that further work is needed to develop a coherent scientific narrative around these new contributions. The current title, for instance, still suggests an investigation of global workspace organization, whereas this is no longer the central question addressed by the manuscript.

Overall, I recognize that the revision represents a substantial improvement and that the authors have addressed several of my original concerns. However, I believe the new direction would benefit from further experimental and conceptual development before publication. I therefore maintain my current assessment, while encouraging the authors to pursue this revised direction in a future submission, giving more time to revise the paper to maximize readability, and guarantee robustness of their approach and results.

Thank you for reading the revised paper and for this careful assessment. We are glad that we addressed several of your concerns and you find the third contribution interesting.

We agree that cancellation does not imply erasure. Section 6.3 shows that the information is still used: when we replace the bridge in loop 2, the final answer breaks, and a wrong bridge breaks it about 27 points more often than a missing one.

What separates rebuilding from restoring a preserved answer is the intervention of Section 6.3. If the later loops only restored what the carried state holds, an answer written into loop 2 would come back as the final answer. It does so on 18.7% of questions in Ouro, against 56.1% in Gemma 3 12B and 74.4% in Llama-3.2-3B after the corresponding edits. The same happens when we write in the correct answer on a question Ouro gets wrong: Ouro mostly does not keep it. So the later loops neither copy the answer they receive nor check whether it is right. They work the answer out again from the rest of the state, and this is what we mean by rebuilding. Section 6.2 shows it without any edit: loops 3 and 4 start with the correct answer already on top, and still go through the tempting answer before arriving at the correct one again. Combining the observations and interventions experiments, we think the rebuilding is the most plausible explanation.

We appreciate the time you gave to both versions of the paper.

Reviewer_XW4WOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates workspace-like representations in looped transformers by extending the Jacobian lens through virtual unrolling and distance-controlled readouts. The authors examine Ouro-2.6B and Huginn-0125, using Qwen3.6-27B as a standard Transformer baseline. Through semantic readouts and causal interventions, they observe different information maintenance patterns: Ouro exhibits evidence consistent with cross-loop semantic reconstruction, while Huginn's information accessibility depends heavily on repeated input reinjection. The paper also examines how intervention timing and scope affect model behavior.

Critical strengths
  • The extension of the Jacobian lens to recurrent architectures is technically motivated and addresses potential readout limitations.

  • The paper includes diverse experiments, particularly causal interventions, providing useful observations about information maintenance across recurrent computation.

Critical weaknesses
  • Regarding §4.1 & §5.1, the failure of cross-loop semantic transport under the mean Jacobian lens does not necessarily imply that Ouro reconstructs semantic information, since nonlinear or input-dependent transformations could still preserve it. Could the authors further validate this interpretation using per-example Jacobians, nonlinear alignment, or cross-loop activation patching?

  • About §6.1, the interventions are applied across all token positions and multiple layers, making it difficult to identify where the causal effects actually originate. It would be interesting to see whether more localized interventions produce similar effects, and whether these findings generalize to more complex multi-step reasoning tasks rather than simple semantic associations?

  • Two different looped transformers consist of two completely different training objectives, including architecture, model size, and training procedures, so its not surprising that we will find differences in their internal workspace. But in fact, i am more curious about how those different training procedures can introduce those differences.

  • I suggest expanding the related work to include causal interpretability research, in both general llm and latent reasoning areas.

Questions for authors

N/A

2 comments

We thank the reviewer for the careful reading and the concrete suggestions. Section, figure and table numbers below are those of the revised paper.

W1. Does the failure of the mean Jacobian across loops imply that Ouro reconstructs? Per-example Jacobians, nonlinear alignment, cross-loop interventions.

We ran the two requested checks on the lens (part 1). They do not settle whether Ouro rebuilds its state, so we tested that directly (part 2).

1. Per-example Jacobians and nonlinear alignment (Section 3.1, Table 1; Appendix B.1). We use two-hop questions whose intermediate concept (the bridge) does not appear in the prompt. We map the state from Ouro's loop-2 band to its loop-3 band and measure how often the bridge ranks first among control concepts of the same category. This rate is 19% for the mean Jacobian, 35% for the Jacobian of each question's own forward pass (Table 9), and 86% for a nonlinear map fitted on ordinary text. Thus, the bridge representation in loop 3 is predictable from the loop-2 state even when the Jacobian readout fails. These checks establish a limitation of the lens; they do not determine whether the next loop preserves or rebuilds the content.

2. Evidence that Ouro rebuilds its state (Section 6). A loop rebuilds its state when the content is already in the state it receives and the loop still computes that content again. Three experiments show that Ouro's loops do this.

  • Cancellation followed by rebuilding (Section 6.1, Figure 5). The bridge and answer are already readable in the state entering each later loop. The first three layers add an update almost opposite to that state (cosine −0.99), largely cancelling it, and the lens readings fall sharply. The remaining layers restore both readings, following a similar layer-by-layer trajectory to loop 1. Qwen shows no such cancellation after cuts at layers 12, 24 or 36.
  • Competing answers reappear in the same order (Section 6.2, Figure 6). On two-hop questions with a tempting wrong answer, loops 3 and 4 begin with the correct answer already at median rank 1. Within each loop, the tempting answer overtakes it before the correct answer returns to the top in the final layers. Later loops therefore repeat the progression seen in loop 1, even when they begin with the correct answer available.
  • Later loops replace an edited answer (Section 6.3, Figure 7). We write a wrong answer into the band of loop 2 and let loops 3 and 4 run without further edits. The edited answer remains the final output on only 18.7% of questions, compared with 56.1% in Gemma 3 12B and 74.4% in Llama-3.2-3B after the corresponding edits to their middle layers.

The two parts give two conclusions. The Jacobian lens does not carry a concept across Ouro's loop boundary, which is why we fit one lens per iteration (Section 3.2). And Ouro does rebuild its state: each loop cancels the state it is handed and writes the content again.

W2. Localised interventions and multi-step tasks.

Localised interventions (Section 5.2, Table 3, Appendix B.3). We restrict the edit by position and by layer. Take "The capital of Canada is the city of". The edit moves the concept Canada onto France and succeeds when the answer becomes Paris. The argument tokens are those of the word the question is about, here "Canada"; the final position is "of", where the answer is read. Successes over 192 trials:

Layers editedEvery positionArgument tokens onlyFinal position only
Ouro, loops 2 to 4738123
Ouro, loop 2 only40430
Ouro, loop 3 only56510
Ouro, loop 4 only481429
Qwen, band80584
Huginn, all three windows25717

We invite the reviewer to look at Figure 4, which shows these results more clearly.

The effect originates at the argument tokens in Ouro's loops 1 to 3 and in Qwen, and moves to the later positions in Ouro's loop 4. In runs with two to six loops, the drop at the argument tokens follows the last loop, whichever number it has. On Huginn only the last window can be edited, and there the edit acts at the final position; 4 to 6 successes is its level of no effect. By layer, the effect is spread over the band: any four neighbouring layers give at least half of the band's effect, and the best single layer gives 7 in loop 2, 28 in loop 4 and 31 on Qwen. An erasure needs every position: removing the concept at the argument tokens alone keeps 40 of 42 answers on Ouro, 34 of 36 on Qwen and all 28 on Huginn.

Multi-hop questions (Sections 5.1 and 6.3, Appendix B.4). The edited concept can be one the prompt never states. In "The language spoken in the country where the Amazon River ends is", the edit moves the unstated Brazil onto Mexico and succeeds when the answer becomes Spanish. On 189 two-hop and 46 three-hop questions, the edit succeeds on 73% of two-hop and 62% of three-hop trials on Ouro (329/449, 52/84), 69% and 64% on Qwen (181/262, 21/33), and 37% and 42% on Huginn (159/432, 38/90). On Ouro and Qwen these rates are close to those of one-step questions (33 and 35 of 48). Each model is edited on the questions it answers correctly, so the rates are not a comparison between models. The scope is narrower than "multi-step reasoning": the unstated concept is always a country, and the result does not extend to arithmetic. On "What is (one plus two) plus one?", the edit of the inner sum succeeds on none of 135 trials on Ouro and on 32 of 138 on Qwen.

W3. How the training procedures produce the differences.

We agree: the paper in its current state does not show how the training procedures produce the differences between the two looped models.

The main reason is that the two models were pre-trained at very different scales: Ouro on 7.7T tokens and Huginn on 0.8T (Table 2, Appendix A). On the two-hop questions of Figure 3, Huginn answers 25 of the 90 items unedited, against 43 for Ouro and 57 for Qwen, and the edit of the unstated concept succeeds on 5, against 26 and 47. A lens can only read and edit what a model computes, so a difference between Huginn and Ouro under the lens cannot be assigned to the way they loop, the training objective or the amount of pre-training (Section 8).

We therefore report what the experiments show. Huginn's answer is readable in a short window before each lens target and can be edited only in the last window (Sections 4.2 and 5.2). Removing or swapping the prompt that Huginn feeds in again at every recurrence, with the recurrent state untouched, shows that it computes its answer again from that input each time (Section 4.2).

For Ouro we could go further, because comparable standard transformers exist: Gemma 3 12B, close to Ouro's four loops in inference cost, and Llama-3.2-3B, close in parameters, both with pre-training of the same order (12T and up to 9T tokens). An answer written into the middle of Gemma or Llama usually stays; one written into Ouro's loop 2 is usually replaced (Section 6.3, Figure 7). Among the models we tested, only Ouro rebuilds its state in this way. The cause is open; the paper names two training runs that would separate the candidates (Section 8). Whether Huginn would show the same rebuild if pre-trained to Ouro's level is a question we cannot answer now.

W4. Related work on causal interpretability.

We have added the following references to the related work (Section 7). For language models in general: amnesic probing (Elazar et al., TACL 9, 2021), interchange interventions (Geiger et al., 2021, arXiv:2106.02997) and causal tracing (Meng et al., NeurIPS 35, 2022). For latent reasoning: three studies that apply interventions to continuous-thought models such as Coconut and CODI (Li et al., 2026, arXiv:2602.08783; Aswal et al., EMNLP 2026; Liang & Pan, 2026, arXiv:2602.00449). If the reviewer thinks other work should be cited, we would be grateful for pointers and will add it.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.