acceptodds
Withdrawn from ICLR 2027

Audit Before You Interpret: Basis Freedom in Trained Neural Operators

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

A trained neural operator can make identical predictions in different internal bases, and those bases yield different interpretations and visualizations of its learned representation. We develop a framework for basis-consistent interpretation and visualization of trained neural operators. Its checkpoint audit applies to both purely data-driven and physics-informed models without requiring a linear decoder. We test candidate basis changes on public checkpoints from several architecture families. Each test applies a compensated reparameterization to the stored weights and includes a paired control expected to change the predictions. These tests allow interpretations to be checked across equivalent representations. For physics-informed cavity operators with low-rank linear decoders, we identify parameter-aligned subspaces using the model's own coefficients and distances between decoded physical fields. We fix these subspaces before consulting reference solutions for validation. The identified asymmetry subspace captures about 97% of the energy in the reference asymmetry component. Projections onto each fixed decoded subspace are independent of its internal basis. Embeddings built from field distances that are the same in every basis recover more reference neighbours than embeddings of standardized field-varimax coordinates. These results support physical interpretation through validated subspaces and visualization through distances that respect the audited basis freedom.

Primary area:
interpretability and explainable AI
Keywords:
Neural operators, representation interpretability, basis invariance, checkpoint auditing, subspace identification, representation visualization, scientific machine learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_1W26OpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper shows that trained neural operators can be reparameterized without changing predictions, so axis-level interpretations may depend on basis choice. It proposes a paired audit (compensated reparameterization plus a control) and tests it on 62 public checkpoints. For cavity and Burgers operators with linear decoders, it identifies parameter-aligned subspaces and validates them against reference solutions.

Critical strengths

The paired control is a simple, useful safeguard that separates true symmetries from inactive features, as shown in the U-Net case.

Critical weaknesses
  1. Limited novelty: basis non-identifiability is well known, and the method combines existing tools (symmetry analysis, SVD, varimax). What is fundamentally new is unclear.

  2. Manual effort: candidate transformations must be derived by hand for each model and location, and full automation is not achieved, limiting scalability.

  3. Unclear practical value: results come from small toy-like operators, mostly confirm known effects, and the full-span interpolant outperforms block-wise prediction in the completion task.

  4. Clarity: The paper is hard to follow. It introduces many custom terms (e.g., field varimax, component recall, certified, admissible) with little intuition, uses heavily compressed sentences with frequent qualifiers, and relies on a ~40-page appendix for key definitions and evidence. Figures 2 and 3 are also overloaded with information. A clearer main text with intuitive explanations would improve accessibility.

Questions for authors

Please see weakness.

Reviewer_aMp5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper studies how prediction-preserving changes of basis can alter the interpretation of neural operator representations. It proposes a checkpoint-level symmetry audit and a field-metric approach for identifying physically meaningful subspaces. Experiments across 62 public checkpoints and physics-informed operators demonstrate basis ambiguities and improved physical alignment and visualization consistency

Critical strengths
  1. The paper addresses an important interpretability problem: internal features can have different meanings under transformations that preserve model predictions.
  2. The experimental evaluation is extensive, covering multiple architectures and including physical validation, reproducibility checks, and confirmatory experiments
Critical weaknesses
  1. Limited novelty and insufficient comparisons with existing interpretability methods. The proposed approach builds on established techniques such as symmetry analysis, SVD, and varimax. While the authors provide comparisons with SVD/POD, they do not evaluate against other interpretability approaches, such as linear probing or sparse dictionary learning. It remains unclear whether the proposed subspace identification method offers substantial advantages over simpler alternatives.

  2. Limited generality and broader significance for ICLR. Given that the paper is submitted under interpretability and explainable AI, I find the broader applicability of the proposed interpretation method insufficiently demonstrated. While the checkpoint audit covers multiple architectures, the main physical subspace identification experiments focus on low-rank neural operators with linear decoders. The authors provide limited evidence that this approach generalizes to more complex neural representations. Together with the missing comparisons, this makes it difficult to assess the broader significance of the contribution for the interpretability community.

Questions for authors
  1. The paper evaluates interpretability through parameter alignment, component recall, and subspace stability. Could the authors clarify why field varimax is expected to outperform simpler approaches such as linear probing or alternative factorization methods on these same metrics?
  2. Can the proposed physical subspace identification method generalize to higher-dimensional representations or nonlinear decoders?
Reviewer_wNZcOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper observes that trained neural operators can make identical predictions in different internal bases, so interpretations of their latent axes and visualisations depend on an arbitrary choice. The authors propose a checkpoint audit that tests which weight reparameterisations leave the output unchanged, each paired with a control that should change it; moreover, for physics-informed operators with linear decoders, they propose a procedure that identifies subspaces associated with each physical parameter and validates them against reference solutions.

Critical strengths

The paper is concerned with the fact that interpretations of learned representations can depend on an arbitrary internal basis. It addresses this issue carefully, with paired controls and transparent reporting of its development history, which is a very welcome approach that increases clarity.

Critical weaknesses
  • Several formal results appear standard, and the physical interpretation is demonstrated only on small, low-rank, two-parameter operators. It is therefore unclear how much the framework adds in practice, or whether it extends to the larger models that motivate it.

  • The paper is very dense and relies heavily on a long appendix and specialised terminology, which makes the main message hard to extract for someone that doesn't work exactly in this field.

Questions for authors
  • Could the authors state concisely what the work adds beyond closely related studies (e.g. Cain, 2026; Godfrey et al., 2022)?

  • Is there evidence that practitioners draw incorrect conclusions from latent axes of neural operators in practice, beyond the example in Figure 1?

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

Related papers

Discussion (0)

Sign in to comment.