acceptodds
Withdrawn from ICLR 2027

Interaction Locality in Spatial Reasoning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Spatial reasoning models must combine location-bound computation with location-invariant structure. These roles are often read off state names, as when a recursive model's high-level state is called global and its low-level state local. However, a name does not fix how far an update's effect travels, and every task defines its own notion of nearby. To address this, we propose interaction locality, which reports the fraction of an intervention's causal effect, at a calibrated perturbation scale, that stays inside a relation declared in advance over the sites a task admits. We analyse HRM and TRM on Maze-Hard, Sudoku Extreme and ARC-AGI, compare them with a matched non-recursive Transformer, and extend the measurement to a 3D grounding model. In the recursive models, a perturbation of the low-level state lands more tightly in the high-level write it produces than in the next low-level call, an ordering that holds for every task-defined relation we test. Under the primary intervention, a size-matched control shows that ordering is propagation in four cells, retention at the perturbed site in one, and unresolved in the cell with the highest raw score. With added depth the non-recursive baseline and HRM keep spreading a perturbation, whereas the weight-tied recursive model plateaus. Because the score is also differentiable, using it as a fine-tuning regularizer raises measured locality without costing accuracy.

Primary area:
interpretability and explainable AI
Keywords:
spatial reasoning, interpretability

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_DyjmOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a method/framework to measure how far the effect of a small change inside a model spreads and compares this against a task-defined notion of "nearby" (maze neighbors, Sudoku boxes, ARC objects, 3D object distance). This is then applied to two recursive reasoning models, namely Hierarchical Reasoning Model (HRM; Wang et al. (2025)) and the Tiny Recursive Model (TRM; Jolicoeur-Martineau (2025)) and finds that changes to the low-level state stay more concentrated in the high-level update than in the next low-level step. The paper reports further depth profiles against a non-recursive baseline, a 3D grounding model (MTU3D), and a fine-tuning loss that increases the score.

Critical strengths

I have to say that I really had hard times reading and comprehending the paper, these LLM-polished papers use a language that is very dense and is somehow difficult to understand. But from what I understood is that the strengths of the paper are:

  1. The paper seems to make a correct and useful point, which is that a "fraction of effect nearby" score always includes the perturbed site itself, so it can look local just because the effect stays where it started. Separating this out changes the conclusion in two of six main cases.

  2. The authors are honest about their own weak spots, reporting calibration failures, off-distribution perturbations, and cases where results reverse.

Critical weaknesses
  1. It seems that the main finding depends on how strongly the model is perturbed. Apparently, it holds with very large noise but often reverses with milder noise.

In essence, I am a bit unsure about the paper but I am leaning rather negative and would argue for reject because the paper’s main finding only holds when the model is hit with unrealistically large noise, and it largely disappears or reverses under milder or more realistic perturbations.

Questions for authors
  1. What I did not understand: Why do Table 1 and Table 2 report different values for the same quantity (e.g., TRM/Maze: +0.230 in Table 1 vs. .143 in Table 2)?
Reviewer_i1VnOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes interaction locality, a measure of the fraction of intervention-induced activation change that falls within a task-defined neighborhood. It applies the measure to HRM and TRM on maze, Sudoku, and ARC-AGI tasks, includes a non-recursive Transformer comparison, and extends the analysis to 3D grounding. The authors introduce controls to distinguish source retention from propagation and use the differentiable locality score as a fine-tuning regularizer.

The definitions and controls are explained clearly, and the appendix is unusually candid about calibration and intervention limitations. The metric and control decomposition are useful, but the demonstrated implications for reasoning quality remain limited. The algebraic definition and source-retention decomposition are consistent. Serious concerns remain about achieved intervention calibration and interpretation of large off-distribution perturbations.

Critical strengths

The paper correctly challenges the assumption that a high-level or low-level state name determines its spatial causal reach. It offers an explicit effect-mass metric, a task-specific neighborhood definition, and size-matched controls. Importantly, it acknowledges that source retention can inflate raw locality and examines alternative perturbations. These methodological disclosures are a substantive strength.

Critical weaknesses

First, the intervention calibration is not reliably achieved in the primary analysis. Appendix H reports that five of twelve self-drop-calibrated analogues have confidence intervals excluding the 30% target. Some achieved self-drop rates differ substantially from the target. The appendix also describes calibration using a ten-sample estimate with unseeded noise, making the original calibration non-reproducible exactly. This weakens comparisons presented as if intervention strength were standardized. The authors' scale sweeps and recalibration analyses help, but do not remove the primary comparability concern. Moreover, several injected perturbations are much larger than natural activation variation, raising the possibility that measured propagation reflects responses to artificial states rather than normal computation.


Second, raw locality can improve through retention at the perturbed site rather than useful task-aligned propagation. The identity L = s + (1-s)L_off makes this explicit, and the authors' own regularization results indicate that most of the increase comes from source retention. Thus the paper establishes that its metric can be optimized in the tested setting without a measured accuracy loss, but not that the optimization improves reasoning, coordination, or robustness. The within-H versus within-L ordering is an empirical finding under the chosen interventions and relations; it should not be interpreted as proof of semantic hierarchical roles. The non-recursive baseline also differs in final competence and does not isolate recurrence from all other architectural and training factors.

Questions for authors

Do the main architecture rankings and within-H versus within-L differences persist after calibration is repeated with fixed seeds and matched achieved self-drop? Can the authors report uncertainty across repeated calibration fits? Do the main results persist under near-manifold interventions of magnitudes comparable to natural activation variation? Can locality regularization increase source-excluded propagation or independently measured reasoning performance, rather than mainly source retention? How sensitive are the conclusions to alternative task-valid neighborhood relations?

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

This paper proposes a measurement called interaction locality to improve the interpretability of spatial reasoning models. The authors use the proposed metric to analyze HRM and TRM models on several grid-based reasoning benchmarks.

Critical strengths
  1. The research problem is interesting.
Critical weaknesses
  1. Poor presentation and readability. The paper is difficult to follow, particularly in terms of understanding the central research question, the proposed methodology, and the key findings. The presentation introduces numerous technical definitions, measurement conventions, and experimental details without sufficiently explaining their motivations or connections. As a result, I find it is very challenging to identify the main contribution and understand what the experiments actually demonstrate.

  2. Unclear significance of the proposed locality metric. While the proposed metric quantifies how perturbations propagate through hidden representations, higher interaction locality does not necessarily indicate better spatial reasoning or more meaningful information exchange. In particular, the experiments show that increased locality can largely result from self-concentration rather than task-relevant propagation. The connection between the proposed metric and actual reasoning capabilities remains insufficiently established.

  3. Limited generalizability of the evaluation. The main experiments on HRM and TRM focus on discrete, grid-structured reasoning tasks. Although the authors include a 3D grounding experiment, it does not directly validate the findings regarding high-level and low-level computation in more general settings. The applicability of the conclusions to continuous spatial reasoning or other hierarchical architectures remains unclear.

Questions for authors

None

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

Related papers

Discussion (0)

Sign in to comment.