acceptodds
Withdrawn from ICLR 2027

Beyond Frame-Correctness: Non-Identification and Actionability of Equivariant Attribution

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Attributions for structured inputs can depend on arbitrary choices of coordinate frame, gauge, or discretization, and a growing literature addresses this by enforcing *frame-correctness* — requiring explanations to transform consistently with the symmetries of the model. We ask what this property establishes beyond soundness. First, equivariance and completeness do not identify a unique attribution rule: for path attributions on a Lie group, we show that left-covariant path families yield frame-correct attributions, construct a second complete left-covariant family on , and give an explicit analytic witness showing that two admissible rules can assign different credit. On a docking scorer, two frame-exact rules select different degrees of freedom on of inputs. Second, we test whether correcting frame dependence or adding finer attribution improves model-external outcomes. In matched intervention studies, repairing IG removes frame-dependent decision flips but does not measurably improve pose quality; adding four exact-Shapley-derived features to a matched reward-model policy changes Success@1 by pp ( CI ). Together, these results separate frame-correctness from rule identification and downstream decision utility.

Primary area:
interpretability and explainable AI
Keywords:
feature attribution, equivariance, explainable AI, actionability, non-identification, native readouts

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_b6TuOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper studies the limitations of some existing attribution axioms. It shows that equivariance and completeness do not uniquely determine an attribution rule. It also evaluates whether frame-correct or finer-grained attributions lead to better downstream decisions. Experiments on the docking and reward models show that removing frame dependence or adding detailed attribution features does not necessarily improve intervention outcomes.

Critical strengths

The paper goes beyond standard attribution evaluation metrics to evaluate attribution through downstream interventions. This is a useful direction for testing whether explanations provide practical value.

Critical weaknesses
  1. The main theoretical contribution is that equivariance and completeness are not sufficient to determine a unique attribution rule. However, the fact that existing axioms are necessary but not sufficient is not new to the community, making the main contribution of this work quite limited. Besides, the paper also acknowledges that non-uniqueness of attribution is already well known.

  2. The link from attribution to downstream action is not rigorously justified. In the docking study, attribution scores are used to select a DOF to modify, while in the reward-model study they are used to choose a restart step. The paper does not rigorously establish that higher attribution should correspond to better interventions in these settings. Therefore, poor downstream outcomes do not clearly show that the attribution itself is unreliable or unactionable.

  3. The paper mainly focuses on revealing the limitations of attribution methods, while offering limited constructive progress. It shows that common attribution properties are insufficient, but it does not provide a better attribution rule or an effective way to connect attribution quality with actionability.

  4. The presentation and organization of this work are difficult to follow. The key contributions and the role of each experiment are not easy to identify quickly.

Questions for authors
  1. What is the new contribution of this work beyond existing findings on the limitations or non-uniqueness of attribution methods?

  2. How do the authors justify that the designed downstream experiments are rigorously correlated with the reliability of an attribution method?

  3. What is an actionable way to use attribution in practice, and how should attribution scores be connected to downstream interventions?

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

This paper exposes the limitations of several desirable mathematical properties of feature attribution methods, focusing particularly on equivariance and completeness. The authors highlight two challenges in established feature attribution methods. First, they demonstrate through a constructive counterexample that equivariance and completeness alone do not determine a unique design of an attribution solution. Their second point questions the utility of feature attribution methods following the two specific guidelines, which is supported by empirical evidence acquired under controlled experiments.

Based on these findings, the authors emphasize the need for further discussion on criteria to guide the design of feature attribution methods, and deeper investigation into the connection between theoretical foundations and downstream utilities.

Critical strengths
  • This paper provides insightful discussions that lead to the exposure of several limitations and open questions in current feature attribution studies.
  • This paper is composed of both analytical and empirical parts, each supporting the respective limitations.

The two strengths provide a solid foundation supporting the arguments that have been made, which are relevant to the field and can have practical implications.

Critical weaknesses
  • Limited contributions and implications. While the authors expose several limitations of existing research, the implications behind these findings remain unclear. Given that the two investigated properties are non-identifying, should we find replacements for them, or expand the property set, and in what way? These more important questions are not discussed, and the paper delivers few concrete suggestions for future improvements.
  • Limited coherence between the arguments. The two exposed issues, non-identification and limited downstream utilities, are presented as two parallel points that are not sufficiently connected.

The limited contributions particularly concern me and drive my recommendation for weak rejection. Overall, this paper reads more like a position paper than a main track paper to me.

Questions for authors

How would explainability research benefit from the insights provided by this paper? How do the two highlighted limitations connect to each other? There are more fundamental properties for feature attribution methods, e.g. sensitivity; why have the authors chosen completeness and equivariance, given that the latter is less popular than many others? What is repaired IG? I failed to find its definition or a detailed description in the main text.

Reviewer_mT82OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper critically re-examines the concept of "frame-correctness" in geometric and post-hoc attribution methods. While recent literature treats frame-correctness as a key benchmark for explanation soundness, the authors investigate what it accomplishes beyond mere mathematical consistency. They present two main theoretical and empirical findings, .i.e, rule non-identification and decision utility. Overall, the paper provides a timely and provocative perspective, arguing that frame-correctness is merely a basic soundness requirement rather than a guarantee of rule uniqueness or decision value.

Critical strengths
  1. This work critically re-examines the concept of "frame-correctness" in geometric and post-hoc attribution methods.
  2. This paper presents two main theoretical and empirical findings, .i.e, rule non-identification and decision utility.
Critical weaknesses
  1. Misalignment between explanation quality and downstream policy heuristics. The paper evaluates attribution "actionability" by tying feature attributions directly to simplified downstream intervention policies (e.g., picking a single degree of freedom or restarting generation at arg min attribution). However, feature attribution methods were originally designed to explain model behavior to humans or identify important input components, rather than serve as direct, automated control policies. A failure in a heuristic, single-step intervention policy does not necessarily imply that the attribution itself lacks explanatory value or utility for human-in-the-loop debugging.
  2. Limited methodological innovation. The primary focus of the paper is analytical and empirical auditing rather than proposing a novel framework or algorithm. The technical constructions used to demonstrate non-identification (such as the piecewise rotate-then-translate path on SE(3)SE(3) or the 2D path witness) serve as counterexamples rather than practical, generalizable algorithmic contributions.
  3. Reliance on standard baseline modifications. The "repaired" or frame-correct algorithms evaluated in the experiments are existing constructions from the literature (e.g., group-path IG or PCA canonicalization). The paper does not propose new, actionable attribution algorithms that overcome the identified non-identification or decision-utility bottlenecks.
  4. Narrow domain and task coverage. The empirical evaluation of decision utility is restricted to two primary applications as protein-ligand docking and step-level math reasoning with a single 7B reward model. Key domains where equivariant models are widely used include 3D molecular property prediction, invariant visual recognition, and physical dynamics simulation. However, these major domains are omitted from the real-world actionability evaluation.
Questions for authors

Please refer to the weaknesses.

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

Related papers

Discussion (0)

Sign in to comment.