acceptodds
Under review as a conference paper at ICLR 2027

A Family of Control Decisions Bypasses the J-Space

Abstract

Reasoning language models carry reportable content in a small, causally privileged subspace of the residual stream, a verbalizable “global workspace” built from that content Gurnee2026. We ask whether this content-derived workspace also carries the model's control decisions: choices like when to stop, call a tool, or refuse. Clamping the workspace to its clean value removes most of the causal effect of a content perturbation that would otherwise pass through it, but leaves a control decision's effect essentially untouched. This holds under the actual sparse-gradient-pursuit decomposition, beyond a linear approximation, and generalizes across model architectures and across a family of control tokens including refusal. Content is broadcast through the workspace; control decisions are computed outside it and surface only through their own readout. The implication we note is that a model can therefore report a control decision faithfully without ever pointing to what caused it.

Primary area:
interpretability and explainable AI
Keywords:
Mechanistic interpretability, global workspace, chain-of-thought monitoring, causal mediation analysis, control tokens, language models

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_rSPkOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper investigates whether the workspace (J-space), previously identified as carrying verbalizable content, also mediates control decisions in language models, including stopping generation, ending reasoning, tool calling, and refusal. The authors construct steering directions associated with control behaviors by computing differences in mean residual-stream activations. They then test whether the effects of these directions on model outputs persist when J-space activations are clamped to their values from the unperturbed run. The authors also use a content-related steering direction as a positive control. Experiments on Qwen3-4B show that clamping J-space substantially reduces the effect of content-related steering (65% mediation), but has almost no effect on steering the stop-generation token. The authors further observe similar patterns across different control tokens and model architectures, and investigate how this phenomenon changes across different stages of post-training. Based on these findings, the authors argue that control-related signals can influence model outputs without passing through the identified content workspace, and discuss the potential implications for chain-of-thought monitoring.

Critical strengths
  1. The research question is clear and relevant. The paper raises a well-defined and worthwhile question: whether the J-space previously identified for verbalizable content also participates in the model's control decisions. Understanding how language models make decisions such as stopping generation, calling tools, and refusing requests is important for both mechanistic interpretability and model transparency research. I believe this is a worthwhile research question to explore.

  2. The experimental evaluation covers a relatively broad range of settings. The authors conduct experiments across multiple model architectures and different types of control decisions. In addition, the paper includes a content-related positive control and performs a series of robustness checks, such as varying the steering strength and using different geometric metrics.

These experiments are valuable because they help determine whether the observed phenomenon is specific to a particular model or experimental setting. However, these experiments alone are not sufficient to fully establish the mechanistic explanation proposed by the authors.

Critical weaknesses
  1. The main experiments do not sufficiently support the paper's central mechanistic conclusion. The paper claims that control decisions are computed outside J-space. However, the main experiments actually test whether the effects produced by artificially constructed steering directions persist after J-space is clamped. For example, the stop direction is constructed by comparing activations at the stopping position in a reasoning trace with those at the midpoint of the thinking block. It is currently unclear whether this direction captures the actual stopping decision or merely reflects other differences between these two positions. More importantly, demonstrating that an artificially injected direction can influence the stop-token logit without passing through J-space does not establish that the model's stopping decision during natural operation is also computed outside J-space. Although the appendix provides some additional ablation evidence, this issue remains insufficiently addressed, particularly for the other types of control decisions.

This is a fundamental concern because the paper's central conclusion relies on the assumption that the results of the steering experiments reflect the model's actual underlying decision-making mechanism.

  1. The effectiveness of the J-space intervention has not been sufficiently validated across different experimental settings. The paper defines the workspace using sparse gradient pursuit, which depends on the current activation state, but most experiments actually use a fixed linear approximation. Section 4.1 examines the actual decomposition, but it remains unclear whether the core clamp-mediation results can be reproduced using the full decomposition method. Furthermore, the content positive control used to validate the effectiveness of the clamp is typically evaluated on a different task distribution from that used in the control-decision experiments. This matters because, in the tool-calling task, the specially constructed content positive control was not clearly validated (mediation of 0.24, compared with 0.33 for the random-direction control). The positive control also failed validation on OLMo-3-Think. Without more thorough validation of the intervention method in the corresponding experimental settings, it is difficult to determine to what extent the reported bypass results can be generalized.
Questions for authors
  1. How do the authors distinguish an artificial J-space bypass produced by steering from the actual computational pathway used for control decisions during the model's natural operation? Specifically, the stop direction is constructed by computing the difference in activations between the stopping position in a reasoning trace and the midpoint of the thinking block. What evidence demonstrates that this direction genuinely captures the computational mechanism underlying the stopping decision, rather than other differences between these two positions, or simply a direction that can directly influence the stop-token readout?

  2. How do the authors validate the effectiveness of the identified J-space across different model architectures and control tasks? Section 3 states that each model uses a Jacobian Lens fitted through layer 34. However, some of the evaluated model architectures, such as Llama-3-8B (32 layers) and GPT-OSS-20B (24 layers), have fewer than 34 layers in total. How exactly is the Jacobian Lens constructed for these models? How do the authors ensure that the resulting workspaces are comparable across different models? Furthermore, most experiments use a fixed linear approximation rather than the activation-dependent sparse-gradient-pursuit decomposition. Do the core clamp-mediation results also hold under the full decomposition method? For experimental settings in which the content positive control was not clearly validated or directly failed validation, such as tool-calling and OLMo-3-Think, how should the authors interpret the corresponding bypass results?

Presentation and Reproducibility Issues:

I noticed several issues related to formatting, experimental reporting, and references that need to be addressed.

  1. Formatting issue (Page 9): The page number overlaps with the text at the end of the Limitations section.

  2. Empty Reproducibility Statement (Page 9): The Reproducibility Statement section is completely empty. Although this section is optional under ICLR guidelines, the authors should either provide relevant reproducibility information or remove the empty heading.

  3. Table formatting: The captions for Tables 1–5 are all placed below the tables, which does not comply with the ICLR 2027 formatting requirements. ICLR 2027 explicitly requires table captions to be placed above the tables.

Bibliographic Inaccuracies:

The reference list also contains numerous inaccuracies in author names and paper titles. By comparing the references in the submission with their original sources, I identified the following discrepancies.

ReferenceInformation in the submissionCorrect informationSource
Darade & Thorat (2026)Authors: Snehal Darade, Sushant ThoratAuthors: Ishita Darade, Sushrut ThoratarXiv:2605.22488
Fadnavis et al. (2026)Third author: Rafael WyssThird author: Felix WyssarXiv:2605.29126
Sun et al. (2026)First author: Chuan-En SunFirst author: Chung-En SunarXiv:2605.09252
Jiang et al. (2026)First author: Xu Jiang; Title: A Value Axis in Reasoning ModelsFirst author: Nick Jiang; Title: The Value Axis: Language Models Encode Whether They're on the Right TrackarXiv:2606.17056
Baherwani et al. (2026)First author: Vasu BaherwaniFirst author: Vatsal BaherwaniarXiv:2607.22925
Mazaheri & Mazaheri (2026)Authors: Parisa Mazaheri, Kian MazaheriAuthors: Parsa Mazaheri, Kasra MazaheriarXiv:2605.20530
Wang et al. (2026)First author: Zhen Wang; Title: New Skills or Sharper Primitives?First author: Zhilin Wang; Full title: New Skills or Sharper Primitives? A Probabilistic Perspective on the Emergence of Reasoning in RLVRarXiv:2602.08281
Wu et al. (2026a)Authors: Jie Wu, Yue Xie, Sheng Lin, Song Zhao, Xin ChenAuthors: Jinman Wu, Yi Xie, Shen Lin, Shiqian Zhao, Xiaofeng ChenarXiv:2603.05773
Wu et al. (2026b)First author: Yue Wu; Title: Tool Calling is Linearly Readable and Steerable in Large Language ModelsFirst author: Zekun Wu; Title: Tool Calling is Linearly Readable and Steerable in Language ModelsarXiv:2605.07990
Galeone et al. (2026)Title: Perfect Detection, Failed ControlFull title: Perfect Detection, Failed Control: The Geometry of Knowing vs. Steering in Language ModelsarXiv:2606.24952
Dutta (2026), Footnote 1Author: S. DuttaAuthor: Chandram DuttaLessWrong article

These errors do not directly invalidate the paper's experimental results. However, such a large number of bibliographic inaccuracies raises concerns about the level of rigor exercised by the authors during the preparation and checking of the manuscript. I recommend that the authors carefully review the entire reference list and correct these errors.

Reviewer_HWHhOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper asks whether the global workspace (the J space) is the causal mechanism underlying a model’s control decisions (stop, tool calling, or refuse). The authors steer the residual streams toward a control decision (e.g., stop) while maintaining the J space coordinate values unchanged as the clean run. They found that maintaining the J space coordinate values did not influence the steering effect on control decisions, but removed most of the influence on the content decisions. They conclude that control decisions are computed outside the J space.

Critical strengths

The topic is interesting and timely. The question of what replies or not relies on the global workspace is of great interest to the mechanistic interpretability community, and can have important implications on safety.

The finding is important, as it raises the possibility that there are structured differences underlying the computations of different types of decisions (content vs. control) that are not captured by existing monitoring tools.

Critical weaknesses

Several design choices could produce the observed results on their own and are not ruled out (see below). The writing is not easy to follow. Many results are reported only as a number without explanation and only with a pointer to an appendix.

Questions for authors

The authors select particular layers to apply the J lens and clamp the coordinate values, and the layers differ across models. These choices are not justified. It is partially reasonable to choose middle-to-late layers, since the global workspace might not have been formed yet in early layers. However, if the layer should be hand-picked for each model, it remains unknown whether the results are an artifact of this particular choice. For example, it could be that content decisions and control decisions enter the global workspace at different layers, and the results only reflect dissociation of layers. To address this, the authors could report the results across all different (combinations of) layers.

The authors state that they use “the content-workspace basis throughout, which excludes any control token’s own readout”. The authors also show that including it changes the result from bypass to routed. This creates an asymmetry between content and control tokens. In the J lens, the dictionary is tied to vocabulary tokens, and for the content decisions, the workspace presumably includes the readout rows of the content themselves (e.g., “China”). In other words, content is allowed to route through its readout direction, but control is not. If this is what explains the bypass result, isn’t the result built in by this design of the clamp basis? The authors argue that it might lead to circularity if they include the control token’s readout, but the same circularity also applies to the content token. To address this, a fair comparison is to also exclude the content token’s readout rows and report the mediation under that condition.

The “Reproducibility Statement” header has no content under it.

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

This paper tests whether LM control decisions (stop, end-of-thinking, tool calls, refusal) route through the "global workspace" subspace or bypass it. Using the clamp test from Gurnee et al., the authors show content drives lose most of their effect when the workspace is clamped (mediation 0.65), but control drives are unaffected (-0.03). The dissociation replicates across seven models and four control-token types. A post-training ladder on OLMo-3 shows SFT installs the bypass, with DPO and RLVR leaving it unchanged. The authors conjecture this limits CoT monitoring's ability to catch a control decision's real driver.

Critical strengths
  • Clean experimental design: the content drive serves as a positive control for the clamp's power, so the control-drive null is a genuine bypass rather than a weak instrument.

  • Breadth of replication is unusual for mech-interp: seven models, four architectures, Bonferroni-corrected battery, multiple dosing conventions.

  • The paper is honest about its failures, OLMo-3-Think's positive control failure, the sign flips in position controls, and the retracted confound are all flagged rather than buried.

Critical weaknesses
  • Primary model is 4B. Cross-model replication is broad but shallow (one number per model). No model at deployment-relevant scale (70B+) is tested, and the workspace was originally characterized on Claude 4.5–4.6.

  • The CoT monitoring implication is a conjecture, not a result. The bridge from "activation-level bypass" to "monitors can be fooled" is prominently featured in the abstract but never empirically tested.

  • Two of four family members (</think>, <tool_call>) produce unexplained sign flips in the position control experiment. The tool-call positive control is also weak (m=0.24 vs 0.33 at N=12). Half the family having unresolved confounds undercuts the "family" framing.

Questions for authors
  • Have you measured the bypass on 70B+ models or models where the workspace was originally found?

  • The sign flip for </think> and <tool_call> in the position sweep. Do you have any hypothesis? Two of four family members failing the position control is a substantial fraction.

  • Have you tried a behavioral experiment, e.g., steering the control drive while monitoring the chain of thought, to test whether verbalized reasoning fails to track the manipulation?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.