acceptodds
Under review as a conference paper at ICLR 2027

Attribution via Distributional Paths for Information Revelation

Abstract

Feature attribution methods often either compare predictions under different subsets of observed feature information, as in SHAP-style methods, or integrate gradients along a path from a baseline to the input, as in Integrated Gradients. We introduce Reveal-IG, which combines these perspectives by defining a continuous path of distributions that progressively reveal information about an input. At each point along this path, the model is evaluated in expectation over partially revealed inputs, and attribution integrates the resulting change in expected prediction. This replaces IG's trajectory through individual input values with a trajectory through information states, while retaining a completeness property: feature attributions sum to the change in expected model output across the path. The framework naturally supports structured perturbations, including variable-resolution image probes and feature-wise uncertainty in tabular data. On synthetic diagnostics, Reveal-IG reduces path artifacts compared to pointwise input-space paths. Across ImageNet classification and tabular regression, Reveal-IG produces stable, signed attributions, leading on sign-sensitive image benchmarks while remaining competitive across standard attribution metrics.

Primary area:
interpretability and explainable AI
Keywords:
Explainable AI, computer vision, shap, integrated gradients

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_iA6sOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper devise Reveal-IG to generate attribution map for the target model. Reveal-IG attributes changes in expected model output along a continuous path of input distributions, moving from broad reference probes toward distributions concentrated around the explained input. Image attribution uses Gaussian probes with changing means and variances; tabular attribution progressively concentrates empirical feature marginals. The method satisfies completeness for the expected responses at its endpoints. Experiments include synthetic diagnostics, ImageNet attribution with ResNet-50 and ViT-B/16, and two tabular regression datasets. Results are particularly strong on sign-sensitive image metrics, while tabular performance is competitive but mixed.

Critical strengths
  • A coherent framework. The distributional formulation provides a common view of continuous information revelation across images and tabular data, with a clear completeness property for the quantity being explained.
  • Encouraging image results. Reveal-IG substantially improves positive-evidence insertion/deletion and Sensitivity-n over the evaluated baselines on both architectures.
  • Useful diagnostics and transparent limitations. The paper includes controlled synthetic examples, endpoint and completeness ablations, confidence intervals, and explicit discussion of runtime and finite-variance endpoints.
Critical weaknesses

(1) The distinction from randomized path attribution is not sufficiently established. SPI (Jeon et al., ICCV 2023) already defines attribution by averaging line integrals over randomized input-space paths. Under Reveal-IG's Gaussian reparameterization, fixing the noise defines an input-space path, and exchanging expectation and integration expresses the image attribution as an expectation of those path attributions. This is a mathematical observation about the proposed formulation, not a claim that the methods are identical: their path distributions and endpoints differ, and the empirical tabular construction has additional scope. However, the related-work distinction between pointwise path averaging and integrating an expected response is insufficient on its own. SPI is neither discussed nor directly compared. The paper should clarify what is conceptually and methodologically new beyond the particular probe family and reveal schedule.

(2) Important practical image baselines are missing. The evaluation includes several relevant IG variants, but omits Grad-CAM and FullGrad. Their different attribution representations do not make them unsuitable for practical insertion/deletion and localization comparisons, particularly on ResNet-50. Completeness is valuable, but does not by itself establish superior empirical usefulness. These comparisons should respect differences in resolution and sign support.

(3) The computational overhead is substantial relative to the demonstrated gains. The paper already reports approximately 1.1 seconds per image versus 165 ms for IG, and approximately 1.5 seconds on Bike Sharing versus 1–3 ms for IG, KernelSHAP, and LIME. The latter overhead is especially concerning because Reveal-IG does not consistently outperform perturbation baselines across tabular metrics. Image settings use 50 integration steps with 10 Monte Carlo samples per step, versus 50 steps or samples for IG and Expected Gradients. Comparisons under equal runtime or model-evaluation budgets, with quality–cost curves, are needed to separate the benefit of the proposed path from additional sampling. The completeness-convergence ablation does not establish attribution-quality gains under matched computation.

(4) The method explains a smoothed prediction rather than necessarily the original prediction. With the default image endpoint σ = .25, completeness concerns the expected output around the input, not f(x) itself. The paper acknowledges this and provides endpoint ablations, but does not sufficiently quantify how often the endpoint changes the target score or class.

(5) Distributional probes do not automatically make intermediate inputs semantically meaningful. Independent Gaussian pixels ignore image structure, while recombining empirical tabular marginals can produce unrealistic combinations of correlated features. These are valid interventional choices, but moving to distribution space does not itself solve the off-distribution problem motivating the method.

(6) Additional evaluation protocols would strengthen the empirical conclusions. Insertion/deletion measures the fixed model's response to perturbed inputs, while RMA measures attribution mass inside object regions. ROAR would provide a complementary test by removing highly ranked features from training and evaluation data and measuring performance after retraining, helping distinguish information removal from sensitivity to distribution shift. The pointing game would test whether the highest-attribution location falls within the target object, complementing RMA's mass-based localization measure. These evaluations would help establish whether the reported gains extend beyond the current protocols; neither should be treated as a complete test of signed attribution faithfulness.

Questions for authors

(1) What is the main difference between Reveal-IG and SPI? Can you compare their results directly?

(2) Can you also compare Reveal-IG with Grad-CAM and FullGrad on the image experiments?

(3) Is Reveal-IG still better when each method is given the same computation time? Can you reduce its cost without losing much accuracy?

(4) How much does adding noise change the model's prediction? Could the better attribution results come mainly from averaging over noisy inputs?

(5) Does the method still work well when input features are strongly related to each other, rather than independent?

(6) Can you add ROAR to check whether removing important features reduces performance after retraining, and the pointing game to check whether the most important pixel lies inside the target object?

Reviewer_8ddaOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper formulates a sort of integrated gradients variant that has a distribution become gradually more concentrated around the input to attribute. Experiments are provided to compare attribution methods.

Critical strengths
  • The idea does have a sort-of heuristic sensibility to it.
  • Experiments and a handful of images are favorable.
Critical weaknesses
  • The integrated gradients method is founded on axiomatic principles. It seems that this paper does not seek to track the axiomatic underpinnings of IG, but instead makes heuristic choices to augment it. What results is a method that has no axiomatic foundations, and we must judge by some hand-picked images and experimental results. Although the experimental results do favor the method, this reviewer believes he has seen many attribution papers with experimental results just like this one, which show that the proposed attribution is superior. Thus, without any theoretical heft, the reviewer is skeptical of the significance of the method.
Questions for authors
  • qθ=Πθqθiq_\theta =\Pi_\theta q_{\theta_i} seems to severely limit the distributions. No spherical shell distributions?
  • IG has an issue of choice of baseline (https://distill.pub/2020/attribution-baselines/). Consider choosing μi=μf=x∗\mu_i=\mu_f=x^*, so the distribution is always centered on the image?
  • For figure 2, ∇f\nabla f does not seem to match up with the image of ff. More explanation would be helpful.
  • Please provide proof of completeness. The claim is not obvious to this reviewer.
  • How does the method handle distributions that attribute pixel values that are out of bounds?
Reviewer_jRVcOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a new feature attribution method, Reveal-IG, which explores modifying path-based attribution methods by making the path move through a sequence of varying distributions

The method is compared to a set of other feature attribution methods such as IG, BlurIG etc. and performs reasonably well

Other experiments are performed examining the nature of the method in more detail

Critical strengths

S1. The reasoning behind the core idea is well motivated and makes intuitive sense. The idea of varying distributions as a path through perturbations seems reasonably novel and promising

S2. Tables 1 and 2 demonstrate positive results for their method in comparison to other methods. I believe this is the strongest set of results in the paper

S3. Several experiments and analyses carried out

Critical weaknesses

W1. The field has become inundated with feature attribution methods, each targeting a different aspect and performing slightly better than other methods in certain settings. While the core idea and goal of this paper makes sense it is still unclear to me why I should use this method over other more algorithms which each seemingly have their own benefits

W2. While text, experiments and figures are plentiful, the text itself was difficult to parse through, and the underlying narrative hard to get to. While results are interesting, I wonder why some of the figures and experiments are relevant to the main research question and if they belong in the Appendix instead eg. Fig 4

W3. Some choices could benefit from more justification. For example, SHAP is compared to RevealIG throughout the paper, but is not covered in 3.2. 3.2 and 3.3 seem to use completely different methods and evaluation protocols but the reason for this is not clear.

W4. Similarly, the rationale behind some strong statements could be justified better. For example, in the introduction, "This discontinuous reveal operation is a severe intervention: a feature jumps to fully observed without allowing evidence to enter the explanation gradually" . Why is this a "severe intervention" ?

Questions for authors

Main questions pertain to the weaknesses. Maybe more specifically:

Q1: For my understanding, what research questions/purpose are Figure 2 and Figure 4 trying to answer? Could the summarized takeaway/s and not just the methodology be cited alongside the figures for more clarity?

Q2: What is the reason for not considering traditional attribution methods like SHAP for images despite it being spoken about throughout the paper? More broadly, why are the setups for 3.2 and 3.3 so different?

Q3. Could the rationale/citation behind the statement in W4 be explained?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.