acceptodds
Under review as a conference paper at ICLR 2027

TameProbe: An Empirical Tameness Spectrum for Neural Decision Boundaries

Abstract

Model-theoretic o-minimality (“tameness”) has recently been proposed as a lens on neural network geometry, but it is a binary structural property that is not directly measurable on a trained model. We turn tameness into an operational, continuous diagnostic spectrum: two cheap, model-agnostic statistics of a network's decision boundary—the oscillation complexity (, sign flips and local extrema of the decision margin along random 1D probes) and the decomposition complexity (, number of distinct linear regions crossed). Across a controlled matrix of MLP architectures, seeds and post-training operators (an alignment-style Jacobian/weight regularizer, knowledge distillation, and 8-bit quantization; measured slices), we find that (i) tameness is not scalar: the alignment proxy lowers both () and (), whereas distillation lowers () while raising (), so the two axes decouple; (ii) 8-bit quantization does not move the spectrum (), a negative result that validates the metrics as capturing function geometry rather than numerical noise; and (iii) on an external adversarial-robustness anchor (), strongly predicts PGD robustness (Spearman , ; the effect of ), connecting our spectrum to the curvature–robustness literature. We further show, on a balanced compositional truth concept across five same-family base/instruct LLM pairs, that both base and post-trained LLMs already sit at the tame end (a nonlinear RBF kernel never beats a linear probe), with post-training leaving this position essentially unchanged. TameProbe gives a reproducible, quantitative vocabulary for how far along the tameness spectrum a model sits and how training operators move it.

Primary area:
learning theory
Keywords:
neural decision boundaries, tame geometry, o-minimality, model diagnostics, adversarial robustness

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

The paper extends o-minimality from a binary property of a function to a continuous metric. The goal is to propose a measurable, continuous version of o-minimality property to quantify the behavior of trained neural networks. The main idea is to check, for each input dimension, how many dimensions change the binary classification decision, and to apply decomposition-complexity metrics from previous work.

Critical strengths
  1. There are a lot of graphics to visualize the proposed metrics and results.
  2. The related work is long and comprehensive. It takes one page.
Critical weaknesses
  1. The paper mentions that few recent papers wrote about "o-minimality". However, it is not clear whether this topic is worth investigating. The paper does not discuss where this "o-minimality" measurement fits relative to other related techniques on evaluating DNN's robustness, so the motivation for the study as written in the Abstract and Introduction is weak. To attract a broader audiences in adversarial machine learning and DNN verification, the paper needs to explain why the proposed metrics, built on "o-minimality," are more useful for proving model robustness and stability than other well established techniques such as Lipschitz-based metrics, DNN verification (either probabilistic or sound and complete), or randomized rounding.

  2. Proposed oscillation complexity (OC) metrics is trivial, because the 1D check is not "sound" and guarantees nothing. Also, the decomposition complexity (DC) metrics: "DC is a boundary-local instance of the region-counting complexity studied by Montúfar et al. (2014); Serra et al. (2018); Hanin & Rolnick (2019)." as authors stated verbatim in the paper, so this is not a contribution of this paper. There is no discussion of the relation between o-minimality theory and the proposed OC/DC metrics.

  3. The proposed method is short, less than one page, which is even shorter than the related work.

  4. It seems like this paper is entirely LLM generated without careful revision by human. I cannot read it smoothly, especially the abstract. There are many signs that the text is generated without careful revision by the author, e.g., excessive usage of ---, constant interruption within every sentence, lots of negative structures like "not" or "rather than" for the LLM to reassure itself that the generated text is correct and it is not hallucinating. A human would write the same paper more directly. Another pattern of recently generated papers by LLMs is that half of the abstract is packed with numerical results, which isn't common before, because traditional human-written papers would concentrate on motivation and high-level concept of the contribution. There are many \textbf{summarize phrase} in thein the paragraph, which is also a sign of an LLM-generated paper.

  5. The language is informal, not suitable to be included in a computer science proceeding, e.g., "how far along a tameness spectrum does this model sit, ... move it, ... they already sit near ...", "Where post-trained LLMs sit", etc.

Questions for authors

I don't have any question, since it is unlikely that I'll change my score.

Reviewer_NBz5OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper empirically characterizes neural decision geometry through oscillation complexity (OC), which counts decision-margin sign changes and local extrema along random input lines, and decomposition complexity (DC), which counts ReLU activation-pattern regions. It examines their responses to regularization, distillation, and quantization, links OC to adversarial robustness, and extends the analysis to other architectures, pretrained vision models, and LLM representation probes. Its central claim is that OC and DC respond differently to training operators, motivating a multidimensional view of empirical tameness.

Critical strengths
  1. Distinguishing margin oscillation from activation-region complexity yields a clear empirical question: do training operators affect them differently? The reported comparisons provide relevant evidence.
  2. The robustness analysis suggests a practical use for OC beyond geometric description, though the scope and added value of the evidence warrant further assessment.
Critical weaknesses

1. The proposed spectrum needs clearer conceptual and mathematical justification

In the Introduction (lines 037–043), you briefly discuss o-minimality, but this does not adequately explain why OC and DC are appropriate coordinates of tameness. Please clarify the conceptual link between this structural property and the proposed statistics.

In Section 3, OC and DC are introduced without sufficient theoretical discussion of their mathematical properties or connection to tameness. For example, under what assumptions is OC in Equation (1) finite, and why are sign changes and local extrema weighted equally? Equation (1) also defines a continuous-domain count, whereas the implementation uses finite samples. Please clarify the treatment of non-strict extrema, constant intervals, and zero-valued margins, and explain how the sampled estimator accounts for potentially missed between-sample oscillations.

In the operator comparisons, you report different responses of OC and DC. However, this alone does not establish their necessity as coordinates of a well-defined tameness spectrum or show that task-specific scalar summaries are inadequate. Likewise, the robustness correlations and model-selection results do not clearly establish the added value of using OC and DC jointly. Please clarify when and why the joint characterization is needed.

2. The architecture-independent claims need clearer scope and justification

In the Introduction, you describe the statistics as model-agnostic, but DC relies on internal ReLU activation patterns and a piecewise-linear region interpretation. Please distinguish OC’s broader applicability from the architectural assumptions required by the joint OC/DC spectrum.

In E1/E2, the experiments focus on MLPs, while E3 adds ReLU CNNs. The later Transformer and pretrained-vision experiments broaden the evidence, but do not establish a common joint OC/DC construction across these families. Please specify which definitions and conclusions about training operators transfer, and under what assumptions.

In Section 4.1, the experiment overview lists E1, E2, and E4 but omits E3 and several later experiments, making the scope difficult to follow. Please update the overview to reflect the full evaluation.

In Section 6, you identify extending DC to convolutional architectures as an open problem, although E3 already uses CNN activation-pattern segments. Please clarify what remains unresolved and how it differs from the construction used in E3.

In Figure 6, the small labels make the cross-architecture results difficult to inspect. Please increase their size for readability.

Questions for authors
  1. In Equation (1), are local extrema strict turning points, all non-strict extremal points, or connected extremal plateaus counted once? How are zero intervals and ties treated? Under which network assumptions is the continuous count finite, and how does the finite-grid estimator approximate that quantity as resolution changes?
  2. What precisely does “model-agnostic” mean separately for OC and DC? Which architecture families support the same joint spectrum, and how should the CNN implementation in E3 be reconciled with the scope statement in Section 6?
  3. Beyond showing that OC and DC sometimes move differently, what evidence establishes the need for the joint spectrum? In particular, what practical assessment benefits from using both coordinates rather than OC alone or another task-specific scalar summary?
Reviewer_vjfiOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper measures network complexity along random lines using what the authors call "oscillation complexity" (OC), which counts sign changes and local extrema of the decision margin, and "decomposition complexity" (DC), which counts ReLU activation patterns. In the main experiment, regularization lowers both measurements, distillation lowers OC and slightly raises DC, and quantization has little effect. OC also correlates with the PGD robust/clean accuracy ratio. Further experiments cover other architectures, real datasets, and LLM representations. The authors interpret the results as a "tameness spectrum" motivated by o-minimality.

Critical strengths

I like the idea of comparing several geometric measurements before and after training interventions. Some examples are quite clear: for d4w512, distillation lowers OC by 17.8% while raising DC by 14.0%. I also appreciate that the authors report real-data results that complicate their main argument. Those deserve more attention. Using forward evaluations to help select robust models could be useful if the result survives the checks I detail below.

Critical weaknesses

The connection to decision boundary tameness is unclear. OC includes margin extrema that may never approach a class change. Small oscillations entirely inside a positive-class region can increase OC without changing the boundary. DC has a related problem: adding ReLU units with cancelling output contributions can increase activation-pattern counts while leaving the margin unchanged. These are potentially useful diagnostics, but they measure different objects, and their disagreement does not establish a quantitative version of o-minimality. Also, the relationship to earlier work is not spelled out. Hanin and Rolnick already study region counts along lines; exact methods such as SplineCam could help validate the sampled estimates on small examples

The statistical unit and measurement protocol need clarification. I could not tell whether the E1/E3 tests use individual probe lines or independently trained models. There are many lines but only five training seeds per architecture. If the tests treat lines as independent evidence about a training procedure, they can understate uncertainty across training runs. Please show paired effects across seeds, especially for the small pooled DC increase.

The robustness result needs stronger controls. The robust/clean accuracy ratio can be misleading. A constant classifier on a balanced task has 50% clean accuracy, 50% robust accuracy, a perfect ratio of one, and zero OC. Table 3 reports complete elimination of OC in several regularized models. That does not establish collapse, but their clean accuracies and prediction distributions need to be shown. The correlation also pools architectures and training regimes. Does OC remain informative within each group, and does it add anything beyond knowing how the model was trained? The model-selection experiment needs the same checks. The speed claim is unverified too. Appendix D gives 800,000 forward evaluations for OC versus roughly 40,000 backward evaluations, plus forward computation, for PGD. Those counts do not establish a tenfold saving; measured runtimes are needed.

The mechanism analysis contains a shared term. The decoupling score subtracts the percentage OC change from the percentage DC change. In the mechanism analysis, all OC changes are negative, andn so the score equals the DC change plus the magnitude of the OC decrease. Correlating this score with the magnitude of the OC decrease therefore correlates a quantity with one of its own components. I don't think that establishes a mechanism. An appropriate null calculation or an independent predictor is needed.

Questions for authors
  1. Which effects survive analysis across training seeds and changes in probe resolution?

  2. What are the clean and absolute robust accuracies of the lowest-OC models? Do the correlations hold within training regime and architecture?

  3. What supports the proposed mechanism after accounting for the shared term in the decoupling score?

  4. Which operator conclusions survive the real-data experiments, and what explains the reversals?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.