acceptodds
Under review as a conference paper at ICLR 2027

The Bound Dominates the Budget: Certified Explanations for Additive Models

Abstract

How stable a minimal explanation of an additive model is depends far more on which sound bound certifies it than on how many features it names. A greedy minimal explainer stops the instant its error budget is met, spending 85.2–98.7% of the tolerance ε under its own unsigned acceptance test; certified against that test the radius is near zero, and buying 0.8 extra features moves it from 0.009 to 0.29 on heart at r_max = 0.5. But that test is not the predicate the explanation makes. The displayed surrogate's error is a sum of signed one-dimensional terms whose worst case over the ball is attained in closed form for main effects (Prop. 4); certified against it, the same minimal explanations have spent 32–38% and already reach 0.34 on heart, and the knob buys a further 1.2×, not 31.7 (at the wider r_max = 2.0 error table, 0.45 and 1.5×, not 14.9). The free change is worth more than the one we sell. Measuring this needs a certificate rather than a sampled sensitivity score, and additive structure makes one cheap: for ŷ = Σ_i f_i(x_i) the L∞ worst-case output change decomposes exactly across features, so a sound per-input radius is d one-dimensional interval problems. We make that practical for continuous neural shape functions — dense-grid extrema padded by a certified analytic operator-norm Lipschitz bound — and pair it with a per-feature worst-case error table, so the certified object is the displayed surrogate: a subset and its functional forms. The same holds one level up: what a backbone leaves after minimality differs up to 11× across model classes unsigned and 1.3–1.4× signed — both off the censoring ceiling, and only on heart is there an order-of-magnitude gap for the bound to close — so the choice of backbone is worth an order of magnitude less than the choice of bound — though which backbone leads is not stable, and EBM keeps heart. We instantiate on VIB-NAM, a bottleneck+bypass additive net, and are explicit that the bottleneck carries none of the results. Scope, costs and the failed tests are stated in §1; a provenance gate recomputes every headline number from the shipped artifacts.

Primary area:
interpretability and explainable AI
Keywords:
certified robustness, explainability, neural additive models, minimal explanations, formal XAI, interpretability, robustness certification, generalized additive models

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_dNy4OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper is about certifying explanations of neural additive models. In particular, minimal explanations are usually more interpretable. The paper seems to argue that the certificate is dominated by the choice of the bounding mechanism.

Critical strengths
  • Obtaining certified explanations are crucial in trustworthy AI.
Critical weaknesses
  • The paper reads very poorly, with typical AI jargon throughout (e.g., "Accuracy is a sanity check, not a claim:", "load-bearing", etc). It does not follow a clear logical thread.
  • The paper also defers many descriptions of their contribution to the (77 page long) appendix, to a degree that makes it impossible to assess the paper in a reasonable time. It is not even clear to me what ist meant with "choice of bound".
  • [1] already finds the minimal explanation for general NAMs in log queries; yet, the paper looks at a simpler function (VIB-NAM), which enabled minor improvements.

[1] Bassan et al. Provably explaining neural additive models. ICLR. 2026

Questions for authors
  • Why do you specifically study VIB-NAM?
Reviewer_vAJpOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The text of my review is entirely human generated. I have not used an LLM for any portion of the reading or analysis of this work nor my write up of my conclusions.

Summary: 

The paper studies certified minimal explanations and their stability. The authors formulate the stability as an optimization problem which changes the criteria of minimality and then demonstrate and explore how this changes the explanations. The paper does not engage with other forms of explanations where sparsity may not be the goal, nor do they engage with existing literature that already certifies stability using optimization-based approaches which have existed for several years [1, 2]. Of course, the problem presented here is novel relative to these works, especially in its focus on additive models, which to the best of my knowledge, is novel. To compute the certified robustness of the minimal explanations the authors use popular methods from neural network certification including CROWN and MIP bounds. The experimental results, are difficult to parse. The networks and datasets that are studied are relatively small on account of the fact that verification does not scale to large networks. This is an inherited limitation and not one distinct to this paper. However, the presentation is not systematically done making it very difficult to fully assess the strengths and weaknesses of the approach.

[1] Wicker, Matthew, et al. "Robust explanation constraints for neural networks." ICLR 2023. [2] Huang, Wei, et al. "Safari: Versatile and efficient evaluations for robustness of interpretability." 2023 IEEE/CVF International Conference on Computer Vision (ICCV). IEEE, 2023.

Critical strengths

The papers focus on certified explanations is a very interesting one, and the problem tackled (additive models) is novel as far as I know. By demonstrating the pitfalls of non-robust explanations in additive models, clearly presenting a problem formulation and establishing baselines, this could mature into a nice paper.

Critical weaknesses

One of the most glaring weaknesses of this paper lies in its presentation. The text reads as being entirely AI-generated which is both difficult to read and understand on a sentence and paragraph level, but more importantly is poorly structured (as in the argument is hard to follow). For example, the abstract immediately — in its second sentence — begins presenting experimental numbers without establishing a problem, method, or baselines. The issue of presentation is persistent throughout the paper making it very difficult to understand exactly what argument is being put forward by the authors. The argument of the paper seems to me to be (1) minimal explanations are not stable which is problematic, (2) we characterize this using methods from neural network verification (3) in a series of experiments the authors demonstrate that they can certify the stability of minimal explanations for additive models. If the authors feel this is not their core argument, I would ask that they help me better understand exactly what their argument is. Throughout the paper myriad points are made about various aspects of the instantiation or modeling choices, but these serve to confuse the main arguments and contributions of the paper. 

Outside of the papers presentation making it very difficult to assess, it does not seem that there is a novel methodology that is presented, only a novel problem setting. The setting is that of certifying the stability of minimal explanations, but the method is to use existing verification tools. I think this serves to limit the impact of the work.

Questions for authors

The main question, about the primary argument of the paper, is contained within the weaknesses section.

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

The paper studies the certified robustness of minimal explanations of additive models and argues that the choice of certification bound plays a larger part in determining it than its size. They use a sound certificate and greedy explainer to show that minimal explanations appear significantly more fragile under the explainer's unsigned bound when compared to a signed bound. This takeway holds across additive model backbones across datasets. The paper also examines models with pairwise interactions, where degraded certification can be partly recovered by displaying interaction terms

Critical strengths
  • Many experiments, ablations, very detailed appendix
  • Honest reporting
  • Simple substantiated, practical main takeaway (to use a signed bound as opposed to an unsigned one to more appropriately allocate the budget)
Critical weaknesses

W1. Paper is very difficult to read, as it is written in a very dense manner and the narratiive/story is very hard to parse. Numerical values are inserted as is again and again within the main text body instead of just summarizing key takeaways. Notation is confusing, for example epsilon on lines 14 and 201 seem to have completey different meanings.

W2. Similar to W1. Some sections are confusing as to why they belong eg. Cost and deep tabular baselines, which probably better belongs in appendix. Similarly it appears if a lot of details and text provided in the main paper can simply be pushed to the appendix

Overall I believe a substantial rewrite is required before publishing

Questions for authors

Q1. Could the authors clarify what they meant by the statement in the limitations: "displaying the pairs restores it, but on wine only with 12 of 20 shown" ?

Table 18 seems to indicate 12 pairs only reaches 61%

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.