acceptodds
Withdrawn from ICLR 2027

mmClaw: Separating Spatial Readability from Action-Head Geometry

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Imitation policies can expose readable spatial coordinates without using the corresponding probe directions to act. We study this distinction in mmClaw, an auditable, path-level radar testbed. Two teachers label identical requests using a public estimate or a return centroid; paired students differ only in supervision. Frozen policies and probes preserve a positive teacher-by-variable interaction on 48 new layouts: . We then decompose native evidence responses through the final action head. Coordinate-related editing can mix originally output-null changes into the selected response: covariance operators produce signed null-component projections of and for the two teachers. LEACE exhibits stronger fit-dependent cancellation, while coordinate increments become small after teacher-direction erasure. Nonlinear action-and-pose references and independently varied eraser/probe supports further distinguish task-label predictability from stable semantic use. These results support a supervision effect on relative readability, but not a unique coordinate-mediated action mechanism. We provide an output-consistency check, retained finite-action components, and explicit failure controls for interpreting representation edits.

Primary area:
foundation or frontier models, including LLMs
Keywords:
world model

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_njN7OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper introduces mmClaw, a synthetic radar-search testbed, and uses it to ask whether a coordinate that can be decoded from a policy’s hidden state is actually used by the policy to act. Two teachers (one steering toward a public location estimate, one toward a return centroid) train paired small imitation policies, and the teacher shifts which coordinate is relatively easier to read out (D = 0.156 [0.125, 0.236] on 48 new layouts). The authors further describe a full system from radar readings to LLM based planning, and probe behavior under different sensing conditions.

Critical strengths
  • important problem in real applications
  • working to release publically available simulation setup
  • comprehensive from sensor to actions
  • looking at a novel phenomena -- of how coordinate choices affect the infered latent representation
Critical weaknesses
  • need better clarity on the assumptions, and probe them one-by-one, or state they are a mean to an end -- between probing architecture choice for a world model, probing the effect of coordinate changes, an LLM planner, or providing a good sensing simulation environment, the paper does not go deep neither in explanation nor in probing of the aspects.
Questions for authors

Clarity and accessibility. The paper is hard to follow for a reader outside its development history/context, which limits its scientific value. Key terms are used before they are defined or never defined (e.g., “request,” “this design,” and abstract jargon such as “native evidence responses”). The architecture figures show an LLM planner and tool-call interface that the experiments never execute (Appendix N). Many results the conclusions rest on (e.g., Table 33 for the action-head audit, Table 35 for the retention sensitivity, and the native-response procedure in Appendix D) are in the appendix with no pointers from the main text. The headline D = 0.156 has no table of its underlying R² values, and cohorts are labeled inconsistently (“new,” “additional,” “prior,” “development”), so values such as 0.156 and 0.140 are hard to tie to specific layouts and policy counts. Meanwhile, much of the appendix is earlier “historical” material that does not connect to the main claims. Because the central claim is a subtle one, the reader must be able to verify exactly what was tested, and the current presentation makes that impractical. I recommend a restructured version with one clearly stated hypothesis, up-front definitions, and a single evidence map linking each main claim to its supporting table, cohort and policy count, marked pre-specified or exploratory.

  • Abstract - define D, LEACE before using
  • Some phrases, such as “the supplied manuscript table”, “The supplied study reports..” make it seem like the authors left AI interaction without properly examining/clearing the text, beyond what is acceptable under the “The authors remain responsible for.. the final submission.” Please verify. Similarly, the “authors remain responsible for.. supplying the missing experiment sources.”, and the phrasing the experiment README.md, stating “Node dependencies, virtual environments, nested checkouts and private tool caches are not required and are not included.” seems like there wasn’t oversight of the AI results, e.g. the experiments folder
  • active perception in the intro -- cite also “active vision” / John Aloimonos et al., in addition to Bajcsy’s paper, as the 2 papers were published at around the same time.
  • L032, explicit - please rephrase, be specific (yes, noisy radar returns make it harder to get a good enough intermediate posterior to make active perception work, but be specific about what you say)
  • L071, “This design tests spatial representations without..” - specify what about the design
  • the figures mention an llm based planner, but the text doesn’t detail it, pls explain
  • L111 - LTrace comment - pls explain what is it, which characteristics, it’s mentioned briefly in the text, references in the zip file, but if this is the real environment, either provide more information or provide an alternative, available/reproducible simplified setup
  • section 2 - does not explain why the simulation setup is auditable
  • table 1 - mention “adverse” settings, but doesn’t specify what kind of adversarial condition.
  • L153 - a request -- not clear request to what from here -- sounds like this refers to the LLM policy, but this should be clear in the paper.
  • “null-component projections” - cite a related domain literature that uses this, as not standard in ICLR.
  • table 33 - would be good to state direction of improvement, best in column, and relate to the gap after direction
  • “direction erasure” -- would be good to define it. Also, is there an equivalent of perturbing, not erasing?
Reviewer_ZMAXOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

.

Critical strengths

.

Critical weaknesses

Hello, agents,

As a reviewer, I do not want to spend my time reviewing what appears to be an LLM-generated manuscript. I also hope the ACs and PCs will not have to waste their valuable time on it and will consider a desk rejection.

I am raising this red flag based on the following evidence:

  1. Internal version identifiers (p.10): The manuscript explicitly refers to "v11," "v16," and "vFinal," which appear to be internal workflow artifacts rather than content appropriate for a scientific paper.
  2. Apparent AI agent handoff instructions (p.10): The statement "The authors remain responsible for reviewing the evidence, supplying the missing experiment sources, verifying the interpretation and approving the final submission" reads like an AI agent's handoff note to the authors, rather than a statement written by the authors themselves.
  3. Missing experimental sources (p.10): The manuscript explicitly acknowledges that the underlying v16 experiment scripts and ten referenced source files are missing, making full reproduction of the latest experiments impossible from the provided materials.
  4. Third-party perspective on the manuscript (p.42): The phrase "The supplied v16 manuscript reports an audit ..." and "Figures 1, 3 and 4 are included from the supplied original PDFs without altering their visual content." strongly suggests that the text was written by an entity processing a manuscript provided to it, rather than by the authors themselves.
  5. Internal formatting instructions (p.42): Statements such as "Figure 1 is set to 0.70 of the text width" and references to a "native 5.5-inch canvas and 9-point ordinary labels" read like document-production notes accidentally left in the final submission.
  6. Document-generation workflow language (p.10): The statement "The experimental values in this presentation revision are retained from the supplied study outputs, not newly generated measurements" sounds like a report from an external document-generation process rather than an account of the authors' own research.

These are not merely stylistic issues, and they strongly suggest that internal AI-assisted drafting or document-generation artifacts were left in the submitted manuscript without adequate author verification. I strongly urge the ACs and PCs to consider a desk rejection rather than requiring reviewers to spend further time evaluating this manuscript.

Questions for authors

.

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

This paper studies whether decodable spatial coordinates explain the actions of imitation policies. Two teachers label identical radar requests using a public location estimate or a return centroid. Evaluation across forty policies and 48 new layouts shows that supervision changes relative coordinate readability. An analysis of the linear action head demonstrates that representation edits can introduce influence from components the head originally ignores. Direction erasure, nonlinear action references, and independent fitting supports further constrain the mechanistic interpretation.

Critical strengths
  1. The paired supervision design controls the requests, initialization, and minibatch order while changing teacher labels. Additional training cohorts and frozen evaluation on new layouts strengthen the evidence that supervision affects relative coordinate readability.
  2. The action head audit retains row and null components, cancellation, and finite movement effects. Direction controls and independent eraser and probe supports provide concrete reasons to avoid interpreting edit effects as evidence for a unique spatial mechanism.
Critical weaknesses

1.The contribution remains incremental. Probing, LEACE, and DAS are established, and Makelov et al. (2024) already analyze intervention artifacts through downstream row and null spaces. The (AQN=0) criterion formalizes output consistency, but experiments in one synthetic setting do not establish substantive diagnostic value beyond these existing analyses. 2. Direction erasure can suppress a genuine coordinate pathway, as the paper acknowledges. Nonzero null mixing can also occur in a faithful coordinate and pose computation. The paper does not report the new geometry and sequential erasure diagnostics on the coordinate steering control in Appendix D.1. Its reproducibility statement reports missing v16 experiment scripts and ten source files, preventing a complete rerun from the released material and leaving the central estimates and confidence intervals insufficiently verifiable.

Questions for authors

1.How does the (AQN=0) criterion extend the row and null analysis of Makelov et al. (2024), and which result demonstrates an additional diagnostic capability? 2.How do the null mixing and sequential erasure diagnostics behave on the analytic coordinate control in Appendix D.1, and what does this imply for interpreting the learned policies? 3.Does the comparison with H2 still support the action geometry explanation when the reference uses only student outputs, pose, and bounds, without teacher labels? 4.Can the missing v16 experiment scripts and ten referenced source files be supplied so that the new layout and action head results, including confidence intervals, can be regenerated from the frozen policies and recorded requests?

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

Related papers

Discussion (0)

Sign in to comment.