Towards Rigorous Explainability by Feature Attribution

A new paper argues that the dominant methods for explaining machine-learning decisions lack rigor and can mislead users, and it charts a path toward verifiable, symbolic alternatives for high-stakes applications [1]. For roughly ten years, non-symbolic techniques have been the default choice when practitioners need to explain complex machine-learning models [1]. These post-hoc explanation tools — often built on statistical sampling or perturbation — produce results that are not formally guaranteed to match the model’s actual reasoning. The paper, authored by Xuanxiang Huang and submitted to arXiv, warns that such methods “can mislead human decision-makers” and calls the absence of rigor “especially problematic” in high-stakes settings [1]. Artificial intelligence, as a field, has cycled through periods of optimism and disappointment since its founding as an academic discipline in 1956, with the current boom fueled by deep learning and the transformer architecture [4]. The explainability gap has grown alongside that boom, as models deployed in medicine, criminal justice, and credit scoring demand justifications that hold up under scrutiny [1][4]. The paper identifies the widespread adoption of Shapley values — and the tool SHAP — as a “prime example of provable lack of rigor” [1]. Shapley values, borrowed from cooperative game theory, assign importance scores to input features, but their application in explainable artificial intelligence often rests on assumptions that do not hold for modern black-box models. Huang’s work surveys ongoing efforts to replace such approximations with symbolic methods that yield verifiable, logic-based explanations [1]. Symbolic AI, which uses formal logic and knowledge representation, was a dominant paradigm in early AI research before neural networks surged after 2012, when graphics processing units began accelerating deep learning [4]. The push for rigorous explainability echoes demands in other data-driven fields. Marketing mix modeling, for instance, relies on statistical causal inference to estimate the impact of advertising tactics on sales, using multivariate regressions and time-series data to produce auditable results [5]. Those techniques were first applied to consumer packaged goods, where manufacturers had access to accurate sales and marketing data, and they have since spread as computing power increased [5]. The paper on symbolic explainability does not propose a commercial tool but instead provides an overview of research directions that could bring a similar level of auditability to feature-attribution in machine learning [1]. Huang’s manuscript was first posted on April 17, 2026, and revised on May 27, 2026 [1]. It does not include experimental benchmarks or user studies; its contribution is a conceptual roadmap for shifting explainable AI from statistically informed guesswork toward methods whose conclusions can be formally verified [1].