Towards Rigorous Explainability by Feature Attribution

Joao Marques-Silva; Olivier L\'etoff\'e; Xuanxiang Huang

Towards Rigorous Explainability by Feature Attribution

T0 review · 2 major / 0 minor · reviewed 2026-07-12 · grok-4.5

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Non-symbolic XAI such as SHAP lacks rigor and can mislead; rigorous symbolic methods offer a sound alternative for relative feature importance.

desk verdict Abstract-only survey/position piece: SHAP-style attribution lacks rigor and symbolic methods are the alternative—known critique, value turns on how well the full text maps the alternatives. read the letter →

arxiv 2604.15898 v2 pith:3BWCD45E submitted 2026-04-17 cs.AI

Olivier Létoffé , Xuanxiang Huang , Joao Marques-Silva This is my paper · ORCID

classification cs.AI

keywords explainableAIXAIShapleyvaluesSHAPfeatureattributionsymbolicmethodsrigorousexplainabilitymachinelearning

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper argues that for roughly a decade the dominant approach to explaining complex machine learning models has been non-symbolic methods that lack formal rigor and can therefore mislead human decision-makers, especially in high-stakes settings. The prime concrete example is the use of Shapley values in explainable AI, most visibly through the SHAP tool. The authors survey and frame ongoing work on rigorous symbolic methods of XAI as a viable alternative, specifically for the task of assigning relative importance to input features. A sympathetic reader cares because explanations that are not guaranteed to be correct can produce false confidence precisely where reliability is most needed; the paper positions symbolic techniques as the route to explanations that can be trusted.

What carries the argument

Rigorous symbolic methods of XAI: formal, logic- or constraint-based techniques that assign relative feature importance with guarantees of correctness, offered as the alternative to non-rigorous Shapley-value approaches.

What would settle it

A high-stakes ML model for which a symbolic XAI method produces a relative feature-importance ranking that is later shown, by formal verification or controlled counter-example, either to be incomplete or to mis-rank a feature that a decision-maker relies on.

Watch

Extended reading notes

Core claim

Non-symbolic XAI methods, exemplified by Shapley-value tools such as SHAP, are not rigorous and can mislead decision-makers; rigorous symbolic methods of XAI constitute a sound alternative for assigning relative feature importance.

Load-bearing premise

That the ongoing symbolic XAI efforts the paper surveys actually deliver practical, rigorous relative feature-importance assignments that avoid the failure modes of SHAP-style methods at useful scale.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Punchline: this is an overview, not a new theorem or algorithm. From the abstract alone it frames a decade of non-symbolic XAI (SHAP/Shapley as the prime example) as lacking rigor and potentially misleading in high-stakes settings, then points to ongoing symbolic methods as the route to rigorous relative feature importance.

What is actually new is modest by design: an organizing survey/position of work already underway in formal/symbolic XAI. Credit where it is earned—the problem statement is clean and useful. The claim that Shapley-value tools can mislead is a real, documented line of critique in this literature, not invented here. Stating that relative feature importance should be assigned with rigor is a fair organizing claim for people who still treat SHAP as the default.

Soft spots are proportional to what we have. We only have the abstract, so coverage, scaling, practical delivery, and independent benchmarks cannot be checked. The reader’s weakest assumption is right: that the surveyed symbolic efforts actually deliver usable rigorous alternatives at useful scale. That is definitional for an abstract-only overview, not a load-bearing flaw inside an argument we can inspect. Novelty is survey-level; circularity risk looks low if the body cites formal results properly rather than only the authors’ prior papers. The stress-test note is correct—there is no equation or protocol here to attack.

Who this is for: researchers already in formal/symbolic XAI, or practitioners who need a map of alternatives to SHAP. A serious editor should send the full paper to referees rather than desk-reject a clear position survey on this topic. I would not bring the abstract alone to reading group, and I would not cite until I see the body. If the full text is a careful map with honest limits, it is useful; if it is mostly self-citation without external comparison, less so. Engage the full paper when it is available.

Referee Report

2 major / 0 minor

Summary. The manuscript is a survey/position paper arguing that non-symbolic XAI methods, particularly Shapley-value approaches such as SHAP, lack formal rigor and can mislead decision-makers, especially in high-stakes settings. It frames ongoing work on rigorous symbolic methods of XAI as a viable alternative for assigning relative feature importance. Only the abstract is available for review; no theorems, proofs, comparative evaluations, or detailed coverage of the surveyed methods appear in the provided text.

Significance. If the full paper delivers a precise technical survey that maps concrete failure modes of SHAP-style methods to specific symbolic alternatives with formal guarantees, it would be a useful orientation piece for the XAI community. The abstract correctly identifies a known line of critique (lack of rigor in popular attribution methods) and points toward symbolic approaches as a remedy. However, significance cannot be fully assessed from the abstract alone, because the practical scope, formal guarantees, and comparative evidence of the surveyed symbolic methods are not shown.

major comments (2)

Only the abstract is available. The central claim that non-symbolic methods (exemplified by SHAP) lack rigor and that symbolic methods provide a rigorous alternative for relative feature importance is stated as framing, but no proofs, counterexamples, formal comparisons, or coverage of the surveyed efforts appear. Without the full text it is impossible to verify whether the claim is supported or whether the practical-delivery gap of symbolic methods is addressed. A complete manuscript is required before a substantive technical assessment can be made.
The abstract asserts a 'provable lack of rigor' for Shapley-value XAI. In a full submission this claim must be backed by at least one concrete, citable formal result or counterexample (e.g., a named theorem or published impossibility result) rather than remaining at the level of overview. Absent that grounding, the load-bearing contrast with symbolic methods cannot be evaluated.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; abstract-only survey framing with no derivation chain to reduce.

full rationale

The available material is only the abstract of a survey/position paper. It states that non-symbolic XAI methods (exemplified by Shapley-value tools such as SHAP) lack rigor and can mislead, and that rigorous symbolic methods are being pursued as an alternative for relative feature importance. There are no equations, fitted parameters, uniqueness theorems, ansätze, or load-bearing self-citations present in the text that could be reduced by construction to their inputs. Self-citation risk cannot be assessed without the full text or bibliography, and the abstract itself does not claim a novel derivation that collapses into a definition or a fit. Under the hard rules, an honest non-finding is required: score 0, empty steps. The residual concern that the surveyed symbolic efforts may lean on the authors' prior work is a coverage/evidence question for the full paper, not circularity exhibited by the given text.

Assumptions & free parameters 0 free parameters · 3 assumptions · 0 invented entities

Abstract-only review of a survey-style paper. No free parameters or invented physical entities appear. Domain assumptions are the standard XAI framing that explanations should not mislead and that rigor matters in high-stakes settings; the ad-hoc-to-paper element is the assertion that symbolic methods are the right alternative path, which the full paper would need to justify.

assumptions (3)

domain assumption Non-symbolic feature-attribution methods (e.g. SHAP/Shapley values) can produce explanations that do not rigorously reflect the model and can mislead users.
Stated as background motivation in the abstract; treated as established rather than re-proved here.
ad hoc to paper Symbolic methods of XAI can assign relative feature importance with formal rigor that non-symbolic methods lack.
This is the paper’s framing of the alternative; the abstract asserts ongoing efforts without exhibiting the formal guarantees.
domain assumption High-stakes uses of ML require rigorous explanations.
Standard normative premise in XAI and AI safety literature, invoked to motivate the survey.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Towards Rigorous Explainability by Feature Attribution." pith.science (2026). https://pith.science/paper/3BWCD45E

@misc{pith2026260415898,
  author       = {Pith},
  title        = {Pith review of: Towards Rigorous Explainability by Feature Attribution},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3BWCD45E}},
  note         = {Machine review of arXiv:2604.15898}
}

read the original abstract

For around a decade, non-symbolic methods have been the option of choice when explaining complex machine learning (ML) models. Unfortunately, such methods lack rigor and can mislead human decision-makers. In high-stakes uses of ML, the lack of rigor is especially problematic. One prime example of provable lack of rigor is the adoption of Shapley values in explainable artificial intelligence (XAI), with the tool SHAP being a ubiquitous example. This paper overviews the ongoing efforts towards using rigorous symbolic methods of XAI as an alternative to non-rigorous non-symbolic approaches, concretely for assigning relative feature importance.