acceptodds
Under review as a conference paper at ICLR 2027

Probing the Emergent Global Workspace of Reasoning Models on Structured Reasoning Tasks

11 views

Abstract

Global workspace theory has become an influential lens on artificial intelligence. We ask whether a workspace-like channel, a limited-capacity state that later computation repeatedly reads and rewrites, arises on its own in reasoning models that were never given one, and if so, where it lives and what it carries. Recursive reasoning models, which refine a latent state rather than extend a token sequence, offer a revealing case, because the state that would have to be broadcast is explicit and re-entrant by construction. We use verifiable, multi-solution mathematical reasoning tasks, where a single intervention can separate whether a model produces a valid solution from which one it produces, and fit Jacobian lenses at corpus, per-question, and per-solution levels to a recursive latent reasoner and an autoregressive LLM. In the recursive reasoner, a subspace amounting to a fraction of a percent of the latent state is causally necessary across most of the trajectory, the recursion repairs single-step damage so that only sustained removal is fatal, and ablating branch-conditioned directions changes which solution is returned while largely sparing whether one is. The autoregressive model instead confines constraint enforcement to a narrow band of late decoder layers, where a single edit destroys validity and identity together. In both, the workspace is a functional role rather than a dedicated module, a thin channel inside the shared state; the recursive model spreads it over reasoning steps while the autoregressive model compresses it into a few adjacent layers, suggesting why recursive refinement can be more robust than autoregressive decoding. The protocol needs only an accessible internal state and a verifier of explicit constraints, and transfers to other models and tasks.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Large Reasoning Models, Mathematical Reasoning, Interpretability, Recursive/Recurrent Reasoning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_CyJHOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work is motivated from global workspace theory and aims to compare the workspace-like state for recursive reasoning models and autoregressive LLM. This work proposes to use multi-solution verifiable problems to probe the global workspace. With multi-solution problems, it is able to observe whether the interupted part influence a local solution or the global answer direction. Following this method, GRAM and Qwen3-32B are used for experiments.

N-queens and edge-coloring are used as the multi-choice task. GRAM reaches relatively high performance and converages on these tasks while Qwen3-8B thinking is relatively bad performances. GRAM is relatively robust to the removal of J-space features. Qwen3-32B mainly be influenced on last few depth.

Critical strengths

The work provides some valuable probing methods and interesting observations. The multi-choice tasks provide a way to distinguish local thinking and global thinking. We might observe more clear reasoning process of models. The comparision of recursive models and autoregressive models is an interesting and important topic. I think this work tries to give some insight from the inner state point of view.

Critical weaknesses

I have some concerns for this work.

  • First of all, the motivation comes from global workspace theory, and the main work is not closely related to this. Jacobian lenes seem to provide low dimension spaces and these spaces are not neccessarily connected to GWT. Is there some better connection between the analysis and GWT.
  • Secondly, The comparison of recurrent models and LLM are not quite fair. The GRAM is a much smaller model and is specifically trained for the eval tasks, while Qwen3-32B is a large model trained on general tasks. For example, I might wonder whether the difference is caused by the difference of specifical task training and general NTP training. The main question is that how should we know that the performance is mainly caused by the model architecture.
  • Thirdly, the multi-choice problems for detecting change of reasoning might not be solid enough, since stochastic generation itself has high converage for solutions (see Fig. 3). Even we do not conduct a removal process, the generation process might change the answer. Fixing the random seed seems not to solve the issue completely. I think a better way might to compare the distribution shift of answers.
  • Finally, I am willing to see a clear and useful formal mechanism is not provided for Interpretation. If it could be done, much more insightful conclusion we could gain from this work.
Questions for authors

See questions in weaknesses.

2 comments

Dear Reviewer CyJH,

Thank you for reading our work and for your constructive feedback. To address your questions and concerns, we ran several additional experiments, including a task-specific autoregressive baseline that isolates the effect of model architecture, and a comparison of the distributional shift of answers on multi-solution problems. We respond to each of your points below.

W1 The connection between the Jacobian-lens analysis and global workspace theory

We agree that the low-dimensional subspaces given by a Jacobian lens are not, by construction, necessarily connected to GWT. The connection is empirical: recent work from Anthropic [1], which largely motivated our study, showed in language models that the space identified by such a lens has several workspace properties, including availability for verbal report, use in internal reasoning, broadcast to downstream computation and limited capacity. For Qwen, we use the lens construction and ablation protocol of [1], and our result is consistent with one of these properties: removing the lens-selected directions impairs reasoning while the model still produces an answer. In our work, we further extend this analysis to a recursive reasoning model, GRAM, where a similar small subspace carries the information the answer depends on. We leave the other properties, such as broadcast, to future work, and we will state this connection explicitly in the revised introduction.

W2 Fairness of the GRAM–Qwen comparison: separating architecture from model size and task-specific training

This concern is well founded and deserves further analysis. To address it, we added a task-specific autoregressive model as a baseline, to isolate the effect of architecture (Appendix B.8, Figure 15). The training set and the intervention method are now both matched: GRAM and this baseline are trained on the same task, the J-lenses are fitted on the same subset, and the J-space construction and removal follow the same schedule, except that the autoregressive model is edited by layer where GRAM is edited by step. This follows from the architecture, since steps in GRAM correspond to layers in an autoregressive model.

Under J-space removal, this task-trained baseline loses validity as Qwen does, while the matched random control stays at baseline. The observation is therefore consistent across the two autoregressive models, one task-trained and small, the other generally trained and large, which indicates that the difference from GRAM is not explained by task-specific training or model size alone.

W3 Whether changes of solution reflect the intervention or sampling randomness, and comparing answer distributions

Thank you for this very helpful suggestion. Accordingly, we added a comparison of answer distributions (Appendix B.5, Figure 12). For each multi-solution puzzle, we compare the distribution of final answers over its branches before and after the removal, using the total variation distance. We control the sampling randomness directly. Each branch is replayed with its stored noise, which reproduces the original answer exactly when nothing is removed, and the matched random control keeps the original solution in nearly all branches. J-space removal moves about half of the probability mass, while a random removal matched in rank and norm moves almost none. So in GRAM the change of solution is caused by the removal and not by sampling.

W4 The lack of a clear formal mechanism for interpretation

We agree. The paper identifies which part of the state matters for the answer, but not yet what that part encodes or what computation the model performs on it. The mechanism we hypothesize is the one suggested by global workspace theory: a small subspace, the J-space, holds the information the answer depends on, and later steps of the recursion read it and rewrite it. The new experiments support parts of this account. The relevant content is low-dimensional (Figure 9), it is rewritten after being removed (Appendices B.6 and B.7), and the choice between valid solutions can be changed by writing one or two directions (Appendix B.4). A formal account would specify what the J-space directions encode and test the remaining workspace properties, broadcast and a capacity limit, which we leave to future work.

[1] Verbalizable representations form a global workspace in language models. Transformer Circuits Thread. https://transformer-circuits.pub/2026/workspace/index.html


Thank you again for your time! We hope we have addressed all your concerns. Should you have any further questions, please feel free to let us know.

Dear Reviewer CyJH,

Thank you again for your constructive feedback. If you feel that our additional experiments and our responses have addressed your concerns, we would greatly appreciate it if you would consider updating your rating. If any questions remain, please feel free to let us know, and we will respond right away.

You noted that comparing recursive and autoregressive models from the point of view of their inner state is an interesting and important topic. We agree, and we hope that multi-solution verifiable tasks, with the fairer comparison your comments led to, give the community a clearer way to look at how these models reason.

Sincerely,

The authors

Reviewer_KRmPOpenReview ↗
Rating
3(+2)
Confidence
3(−1)
Informativeness
3(+1)
Presentation
2
Soundness
2(+1)
Read review
Paper summary

This paper investigates which directions within reasoning models influence the validity of an answer and the specific identity of that answer, using constraint-satisfaction tasks that admit multiple valid solutions and can be verified by a deterministic verifier. The authors construct "Jacobian lenses" for both the Generative Recursive Reasoning Model (GRAM) and the autoregressive model Qwen3-32B, fitting sensitivity directions across three conditioning levels: corpus-level, per-question, and per-solution. Following interventions, the outputs are categorized as the same solution, a different valid solution, or an invalid output. In the analyzed GRAM configurations, persistently removing identified directions significantly reduces the validity rate, whereas single-step per-solution interventions primarily alter the answer identity with minimal disruption to any-solution validity. Experiments with Qwen on the nn-queens completion task reveal heightened sensitivity in the later decoder blocks. Based on these findings, the authors discuss potential "global workspace" functions and the distinctions between recursive refinement and autoregressive decoding.

To be honest, somewhat difficult to follow—perhaps because I am not particularly adept at reading something might to be AI-generated.

*Updated at 18/11/2026: Solved and will further comment below.

Critical strengths
  1. The experimental details are handled with rigor, and the limitations are disclosed transparently. The authors verify that a step-by-step "replay" accurately reproduces the original (unintervened) output logits; they employ paired underlying standard-normal draws and utilize puzzle-cluster bootstrapping—ensuring that multiple branches stemming from the same problem are not treated as entirely independent samples. Furthermore, the authors proactively disclose variations in training configurations, distinctions between historical and new experiments, missing controls, and discrepancies in computational budget allocation.

*Updated at 18/11/2026: Will further comment below.

Critical weaknesses
  1. The key interventions fail to simultaneously account for subspace rank, the realized perturbation norm, and the number of intervention applications, which limits the interpretability of the findings regarding the model architecture. While the random-subspace controls for GRAM are matched for rank, the study does not fully report controls that also account for the realized perturbation norm. Since sensitivity-selected subspaces and random subspaces may account for different proportions of the total activation energy, matching by rank alone is insufficient to rule out the influence of the perturbation magnitude. The situation for Qwen is different: the removed subspace is spanned by ten lens-derived token directions, whereas the control condition uses a single isotropic random direction scaled to the same norm—meaning the perturbation norms match, but the ranks do not. Furthermore, the Qwen experiments concerning per-question and per-solution conditions lack corresponding random controls. More critically, the intervention scope differs. In GRAM, a "single-step" intervention involves editing the model state once at the end of a single supervision step; in contrast, Qwen’s "single-block" intervention involves repeated edits across all prompt prefill positions and every subsequently generated token—it is not a comparable, one-off perturbation. Additionally, the two approaches employ different Jacobian operators and fill rules. While the authors acknowledge that the comparison is qualitative, these discrepancies fundamentally limit the proposed mechanistic explanation—that "recursive models can recover, whereas autoregressive models cannot."

  2. The phenomenon of behavioral recovery is insufficient to definitively identify mechanisms such as "active repair" or the "separation of solution-selection and solution-maintenance channels." Persistent interventions applied from various start steps (up to step 15) simultaneously alter the intervention onset, the number of edits, and the relationship between the fixed basis and subsequent states. For instance, an intervention starting at step 0 involves 16 edits, whereas one starting at step 15 involves only a single edit; consequently, these curves cannot isolate the importance of individual steps. While the observation that single-step interventions have minimal impact—while persistent interventions have a significant effect—is consistent with a recovery process, it could also be explained by representational redundancy, subspace drift, or the accumulation of perturbations. The authors acknowledge in the appendix that they have not directly observed the reconstruction of the removed features; therefore, the stronger claims regarding "active repair" made in the main text should be interpreted with the same caveats.

*Updated at 18/11/2026: Solved and will further comment below.

Questions for authors
  1. Could you provide a comparison between GRAM and Qwen that matches subspace rank and realized perturbation norm? Furthermore, could you distinguish between editing a single token position and applying persistent edits throughout prompt prefill and subsequent decoding in Qwen? If the observed effects persist under these aligned conditions, it would significantly strengthen the claims regarding subspace specificity and architectural interpretability.

  2. Why is the sensitivity in the second GRAM training recipe so much lower than that of the primary configuration? Which phenomena can be consistently reproduced when using independent training seeds with a fixed training configuration, or when testing on samples that were not involved in the selection or fitting processes?

  3. Is there any direct measurement of whether removed components are subsequently restored? Alternatively, have you conducted experiments using fixed-length intervention windows to help disentangle effects such as active reconstruction, temporal positioning, subspace drift, and cumulative perturbation?

  4. Where exactly are the "solution-identity transfer" results mentioned on page 8? Please provide details on the methodology, the solution-identity transfer success rate, the results from random baselines, and the extent to which the validity of the output is preserved. If this experiment does not exist, does it weaken the claim that the channels for "selection" and "maintenance" are distinct?

  5. Would the authors be open to reframing the interpretations concerning the "global workspace" and "architectural robustness" as hypotheses, while focusing the verified contributions specifically on the phenomena of sensitivity and answer identity? If the stronger mechanistic claims are to be retained, which results serve to rule out the alternative explanation that these are simply general output-sensitive subspaces?

*Updated at 18/11/2026: Solved and will further comment below.

6 comments

Dear Reviewer KRmP,

Thank you for taking the time to review our work, and for your thoughtful and constructive feedback. We have revised the paper and added several experiments in response. Our replies to each point follow.


W1 Interventions and controls are not matched in rank, perturbation norm, number of applications, scope, or Jacobian operator and fill rule, which limits the architectural explanation

Thank you for this detailed analysis. We have added several experiments to address each point.

Rank and norm. For both GRAM and Qwen, we reran every condition with random controls matched in both rank and realized norm, including the per-question and per-solution conditions (see the revised Figures 4 and 5, and Figure 9 for GRAM at k = 10, 16 and 64). Overall, the random control stays at baseline and has little effect on the solution.

Intervention scope. As mentioned in our response to Reviewer YQCf, the intervention procedures are deliberately designed for each architecture. GRAM has no token sequence, so its state is edited between recursion steps. Qwen's state is spread over token positions, so an edit at a layer is applied at each position.

To address your concern more directly, we added several experiments. Figure 21 presents three schedules for the Qwen intervention: single-token removal, single-layer removal (corresponding to single-step removal in GRAM) and cumulative removal (corresponding to cumulative removal in GRAM).

We also added a task-specific autoregressive model to isolate the effect of architecture (Appendix B.8, Figure 15). The training set and the intervention method are now both matched: GRAM and this baseline are trained on the same task, the J-lenses are fitted on the same subset, and the J-space construction and removal follow the same schedule, except that the autoregressive model is edited by layer where GRAM is edited by step. This again follows from the architecture, since steps in GRAM correspond to layers in an autoregressive model.

W2 Behavioral recovery does not identify "active repair" or separate selection and maintenance channels; persistent interventions confound onset with number of edits

Thank you for this careful point. We agree that behavioral recovery alone does not identify a mechanism, and we added experiments that look at the state directly.

In single-step removal, validity is almost unchanged in all three conditions (Figure 10), where each start step receives exactly one edit. We then examine the state after this single removal. Figure 13 shows the per-solution study: in branches that keep their solution, the removed component is restored within a few steps of the intervention (0.61–0.68 one step later, all of it by the final step). Because this lens is fitted per trajectory, a portion of branches switch to another valid solution, and in those the component is not restored. We expect the same-solution restoration to hold for the corpus and per-question lenses as well, and we are still running these experiments. The sliding-window removal shows the same restoration (Figure 14(c)) at later steps, with removal windows of 1, 2, 4 and 8 steps.

Specifically, on the alternatives you list: Redundancy: a spare copy elsewhere would leave the rest of the state undisturbed and the removed directions empty. The removed directions do not stay empty while the answer is produced elsewhere; the component returns, with cosine 0.96–1.00 to the unedited one. Accumulation of perturbations: a random subspace matched in rank and norm, removed on the same persistent schedule, leaves validity at baseline (Figures 4 and 9). Subspace drift: our recovery claim no longer rests on comparing single-step with persistent removal. Figure 13 observes the recovery directly: after one removal, the state returns to the unedited trajectory both in the removed directions and in the J-space of each later step. Figure 14(c–e) further shows that after the same start-step basis has been removed for up to eight steps, the removed part regrows, so these directions are still in use at later steps.


We will further answer your questions in the next comment.

Q1 A GRAM–Qwen comparison matched in rank and norm, and single-token versus persistent edits in Qwen

Yes, as mentioned previously, we have reran every condition with random controls matched in both rank and realized norm, and updated Figure 4 and 5 accordingly. Figure 21 presents three schedules for the Qwen intervention: single-token removal (the last prompt token), single-layer removal, and cumulative removal. We note that single-token removal in Qwen has no exact counterpart in GRAM, since GRAM has no token sequence and a single-step removal there edits the whole state at once. Our every-position schedule for single-layer removal and cumulative removal follow Anthropic's work [1], which does not report a single-position version of the ablation; the single-token condition is our addition in response to your question.

Q2 The lower sensitivity of the second training recipe, and reproducibility across training seeds and on held-out samples

The second training recipe in the previous version used a slightly different training configuration (especially the learning rate scheduler). When we take a later checkpoint from that recipe, its behaviour approaches what we observed for the first recipe. We interpret the lower sensitivity as the model not yet having stabilized at the earlier checkpoint.

We have also run the same training configuration with three different seeds (Figure 7, Appendix B.1). The loss of validity under corpus J-space removal appears in all three seeds, with the matched random control at baseline. Finally, we fitted the J-lens on training puzzles that are disjoint from the evaluated test puzzles (Figure 8, Appendix B.1). The effect on the test puzzles is similar to that obtained with the original lens.

Q3 Direct measurement of restoration of removed components, and fixed-length intervention windows

Yes, as mentioned in our response to W2, we added experiments to examine state recovery and sliding windows. First,

After a one-time removal in the per-solution experiment (Appendix B.6, Figure 13), in branches that keep their solution, 0.61–0.68 of the removed component is restored one step later and all of it by the final step, and the rest of the state returns to the unedited trajectory. In branches that switch, it is not restored: the state converges to that of an unedited branch with the new solution.

The same pattern holds for longer removal windows (Appendix B.7, Figure 14), and it shows what is regrown. At the final step, the cosine between the regrown part and the unedited one is 0.96–1.00 in branches that keep their solution and −0.03 to 0.17 in branches that switch. So the regrown component is specific to the solution the branch ends on. This is expected, because the per-solution lens is fitted on the branch's own trajectory. We expect the same restoration under the corpus and per-question lenses and will add those experiments.

Q3 Direct measurement of restoration of removed components, and fixed-length intervention windows

Yes to both. We added the two experiments you suggest. (Please also see our response to W2.)

Direct measurement (Appendix B.6, Figure 13). After a one-time removal, we track the removed component at every later step. In branches that keep their solution, 0.61–0.68 of it is restored one step later and all of it by the final step, and the rest of the state returns to the unedited trajectory.

Fixed-length windows (Appendix B.7, Figure 14). We remove the per-solution J-space over windows of 1, 2, 4 and 8 consecutive steps, for every start step. This separates the effects you list:

  • Temporal positioning: the window's position matters. With short windows (1–2 steps), only early windows change the answer. With long windows (8 steps), later windows move more branches off their original solution than early ones, and these additional branches mostly become invalid, whereas early windows mostly redirect to another valid solution.
  • Cumulative perturbation: for a fixed start step, longer windows move more branches off their solution, while the rank- and norm-matched random control at the longest window stays at baseline.
  • Active reconstruction: after the window ends, the removed part regrows, and at the final step it matches the unedited one in branches that keep their solution (cosine 0.96–1.00) but not in branches that switch (−0.03 to 0.17).
  • Subspace drift: drift can only matter when directions fitted at one step are used for removal at later steps. In Figure 13 the removal is performed once, at the step where the directions were fitted, so the recovery observed there cannot be an effect of drift. In Figure 14 the regrowth is still observed after the same directions have been removed for up to 8 steps.

[1] Verbalizable representations form a global workspace in language models. Transformer Circuits Thread. https://transformer-circuits.pub/2026/workspace/index.html

Q4 Location, methodology and results of the solution-identity transfer experiment

Thank you for catching this. The "solution-identity transfer" results are now in Appendix B.4.

Methodology. For multi-solution test puzzles, we take two branches that reach different valid solutions. At a single step, we add the leading SVD components of the difference between the two states to the host branch, replay the host's own noise, and record whether it outputs the donor's solution, its own, a third valid solution or an invalid board. The random baseline writes the same singular values along random orthonormal directions, so it matches both rank and norm.

Results. From the early-middle steps onward (step 3, 7), writing only one or two components transfers the donor's solution in most cases, while at the first step the transfer is partial. The random baseline rarely transfers. Validity is preserved: almost all outputs remain valid boards, and few reach a third solution.

Selection versus maintenance. The part of the state difference that lies inside the host's J-space directions carries only a very small share of its norm, and writing that part alone transfers the solution much less often than writing the leading components. We interpret this as evidence that the directions that select a solution largely differ from the lens-selected directions that maintain it, and we present this distinction as a hypothesis.

Q5 Reframing the global-workspace and architectural-robustness interpretations as hypotheses, and ruling out general output-sensitive subspaces

Yes, we would be happy to reframe the paper more around its verified contribution: Lens-identified low-dimensional subspaces exist whose removal dissociates answer validity from answer identity. Global workspace theory motivates the probe, and the architectural robustness is now better supported by the comparison with the task-trained autoregressive baseline (Appendix B.8).

We do want to emphasize that these subspaces are more specific than a generic output-sensitive one. This work is largely motivated by recent work from Anthropic [1]. For Qwen, the connection to global workspace theory is direct: we use the Jacobian lens and the ablation protocol of [1], which showed in other language models that the lens-identified space has several workspace properties. Our Qwen result is consistent with one of them. For the recursive reasoning model GRAM, we observe a similar pattern: a small subspace carries the information the answer depends on. We leave the other properties, such as broadcast, to future work.

[1] Verbalizable representations form a global workspace in language models. Transformer Circuits Thread. https://transformer-circuits.pub/2026/workspace/index.html

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

The paper asks whether reasoning models build up a shared internal workspace that holds the information they need to solve a problem. The authors compare two very different models: GRAM, which repeatedly updates a latent state, and Qwen3-32B, which reasons by generating tokens one at a time. They test both on n-queens completion and edge-coloring of cubic graphs. Their method is to erase specific information from the model's internal states and see what comes out: the same answer, a different but still valid answer, or an invalid one. On n-queens, a single intervention usually doesn't break GRAM. It still finds a valid solution, though sometimes a different one from before. Repeated interventions do much more damage. Qwen is especially sensitive to interventions in a few late layers. The authors take this as a sign that the two models store and use reasoning information in different ways, but they admit the intervention setups for the two models aren't directly comparable. In the edge-coloring experiments, Qwen shows a different sensitivity pattern, so the results seem to depend on the task too.

Critical strengths

Overall, the paper is well written and the related work is clear. I appreciate how the authors identify gaps in our understanding of existing reasoning architectures, particularly what their internal states carry despite strong task performance.

  1. The use of multi-solution problems, including n-queens completion and cubic-graph edge coloring, separates changing an answer from losing the ability to solve the problem. This provides more insight than accuracy into what the model’s internal state supports.
  2. In GRAM one-time removal can redirect the model to another valid solution whereas repeated removal is more disruptive. Paired randomness and random-subspace controls strengthen the evidence that selected internal directions influence solution identity and correctness differently.
Critical weaknesses
  1. The experiments identify internal directions that affect answer validity and identity, but it is unclear what distinguishes these from other representations important for producing an answer. The paper invokes competition, information broadcast, and limited capacity without clearly demonstrating these properties. In particular, the number of removed directions is chosen for the intervention rather than measured as the workspace’s capacity. The findings are interesting, but do not yet establish the stronger global-workspace interpretation

  2. The intervention procedures are not directly matched: Qwen is edited at every processed token position within a selected layer, while GRAM is tested with both one-time and repeated interventions. This complicates the comparison of their ability to recover from disruption. Qwen’s control also matches perturbation magnitude but not the number of removed directions, and its conditioned experiments lack random controls. Although the authors acknowledge these limitations, they weaken the proposed architectural explanation.

  3. Although the paper is generally well written, I found it difficult to track the differences in intervention size, evaluation populations, and application schedules across experiments. A single table comparing these choices would help. There are also unresolved citation keys in lines 120–124 that should be corrected.

Questions for authors
  1. Qwen is edited at every processed token position within a selected layer, whereas GRAM is tested with both one-time and repeated interventions. What evidence separates an architectural difference in recovery from the effects of these different intervention schedules?
  2. Qwen’s corpus-level control matches perturbation magnitude but not the number of removed directions, and the conditioned experiments lack random controls. Could the authors provide a robustness check using controls matched on both properties to check if the effects are specific to the selected directions?
  3. What evidence shows that GRAM reconstructs the removed information after a one-time intervention, rather than finding a valid answer using information retained elsewhere or supplied again by the input? How can these explanations be distinguished?
1 comment

Dear Reviewer YQCf,

Thank you for taking the time to review our work, and for your thoughtful and constructive feedback. We have revised the paper and added several experiments in response. Our replies to each point follow.


W1 The global-workspace interpretation and the choice of the number of removed directions

We agree that this is not a complete account for the global-workspace interpretation. We would like to reframe this work as: lens-identified low-dimensional subspaces exist whose removal dissociates answer validity from answer identity. GWT motivates the probe, and the paper tests one necessary condition. In addition, we ran experiments to test different numbers of removed directions (Figure 9), which serve as a preliminary probe on the occupancy of the workspace. We also conduct state recovery experiments in Appendix B.6 (Figure 13), which partially demonstrate the recursive restoration effect.

W2 The intervention procedures and random controls

The intervention procedures are deliberately designed for each architecture. GRAM has no token sequence, so its state is edited between recursion steps. Qwen's state is spread over token positions, so an edit at a layer is applied at each position. We added several experiments to strengthen the probe. Figure 21 presents three schedules for the Qwen intervention: single-token removal, single-layer removal, and cumulative removal. For both GRAM and Qwen, we reran every condition with random controls matched in both rank and norm (see the revised Figures 4 and 5). In every condition and for every k, the random control stays at baseline, presenting no effect in solution perturbation.

W3 A summary table of experimental settings, and unresolved citation keys

Thank you very much for pointing this out. We summarize the current settings in Table 2 (Appendix A.5), and we are running further experiments to align the remaining protocols. The citation keys are fixed in the revised draft.

Q1 Separating architectural differences in recovery from the effects of intervention schedules

To isolate the effect of the architectural difference, we also studied a small task-trained autoregressive baseline (Figure 15, Appendix B.8). The training set and the intervention method are now both matched: GRAM and this baseline are trained on the same task, and the J-space removal follows the same schedule. Combining the observations (GRAM single-step removal in Figure 10, cumulative removal in Figure 4, and autoregressive single-layer removal in Figure 15), we can partially conclude that the recursive architecture plays an important role in the recovery process.

Q2 Random controls matched in both rank and norm

(See also our response to W2) Thank you for pointing this out. We have added random controls matched in both rank and norm for all experiments, and updated the results in the revised version (Figures 4 and 5). Overall, the random controls have little effect on validity.

Q3 Evidence for restoration of the removed component, as opposed to information retained elsewhere or re-supplied by the input

This is a very good quesiton. We ran two additional experiments that track the removed component directly.

After a one-time removal in the per-solution experiment (Appendix B.6, Figure 13), in branches that keep their solution, 0.61–0.68 of the removed component is restored one step later and all of it by the final step, and the rest of the state returns to the unedited trajectory. In branches that switch, it is not restored: the state converges to that of an unedited branch with the new solution.

The same pattern holds for longer removal windows (Appendix B.7, Figure 14), and it shows what is regrown. At the final step, the cosine between the regrown part and the unedited one is 0.96–1.00 in branches that keep their solution and −0.03 to 0.17 in branches that switch. So the regrown component is specific to the solution the branch ends on. This is expected, because the per-solution lens is fitted on the branch's own trajectory. We expect the same restoration under the corpus and per-question lenses and will add those experiments.

This bears on the two alternatives you raise. The component is restored, so the model is not simply answering from information held elsewhere while the removed directions stay empty. And the re-injected input cannot be the sole source: the input is a partial board, the target is never given at inference, and the puzzles analysed each admit several valid completions, so the input determines the set of valid answers but not which one is reached. A component that matches the branch's own solution must therefore draw on the remaining state or on the stored noise draws.


Thank you again for your time and your careful reading. We hope these replies address your concerns, and we would be glad to answer any further questions.

Comments on the paper

1 comment

Dear Reviewers and Area Chair,

Thank you all for your time and your constructive feedback.

We are glad that all three reviewers recognized the core contributions of this work. The paper "identif[ies] gaps in our understanding of existing reasoning architectures" (YQCf), and its multi-solution design "provides more insight than accuracy into what the model's internal state supports" (YQCf), offering "valuable probing methods" (CyJH) on "an interesting and important topic" (CyJH). The reviewers also confirmed the quality of the evidence: "the experimental details are handled with rigor, and the limitations are disclosed transparently" (KRmP), and our paired controls "strengthen the evidence" (YQCf).

The reviewers' concerns were specific and actionable, and addressing them has made the paper substantially stronger. Below we summarize the new experiments, how they address these concerns, and what has changed in the draft.

New experiments

  • Matched random controls: every removal in GRAM and Qwen is now paired with a random subspace matched in both rank and realized norm (Figures 4, 5, 9 and 18).
  • Intervention schedules: single-step removal in GRAM at all three lens levels (Figure 10), sliding removal windows of 1 to 8 steps (Figure 14), and three schedules for Qwen, namely single-token, single-layer and cumulative removal (Figure 21).
  • Direct measurement of recovery: we track the removed component after a one-time removal (Appendix B.6, Figure 13).
  • Task-trained autoregressive baseline: a model trained on the same data as GRAM, to separate architecture from model size and task-specific training (Appendix B.8, Figure 15).
  • Answer-distribution shift: a comparison of answer distributions on multi-solution puzzles before and after removal (Appendix B.5, Figure 12).
  • Solution-identity transfer: with random baselines (Appendix B.4, Figure 11).
  • Robustness: three training seeds, a lens fitted on training puzzles and evaluated on test puzzles, and ranks k = 10, 16 and 64 (Figures 7, 8 and 9).

How these address the main concerns

  • Controls and matched protocols: validity under the matched controls stays at baseline, and the schedules for the two models are now stated and compared explicitly.
  • Recovery versus alternative explanations: restoration of the removed component is now observed directly, and is no longer inferred from comparing single-step with persistent removal.
  • Fairness of the architectural comparison: the autoregressive baseline matches GRAM in training data and lens construction.
  • Sampling randomness: runs are paired on stored noise, and the effect is also shown as a shift in the distribution of answers.
  • Global-workspace interpretation: we agree with the reviewers that the paper tests one necessary condition. We frame the verified contribution as lens-identified low-dimensional subspaces whose removal dissociates answer validity from answer identity. The workspace reading is a motivating hypothesis, and broadcast and capacity are left to future work.

Revisions to the draft

  • Appendices B.2 to B.8 and C.1 are new, with Figures 7 to 15, 18 and 21.
  • Figures 4 and 5 are updated with the matched controls.
  • Table 2 (Appendix A.5) summarizes the settings of all experiments.
  • Unresolved citation keys are fixed.

We are grateful for the discussion and would be happy to answer any further questions.

Sincerely,

The authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.