Abstracting General Syntax for XAI after Decomposing Explanation Sub-Components
Springer AI, 2024 · University of Florida
Stephen Wormald
Matheus Kunzler Maldaner
Kristian O'Connor
Olivia P. Dizon-Paradis
Damon L. Woodard
What is the Qi-Framework?
The Qi-Framework is a mathematical syntax for explainable AI that writes any explanation as E{i,u} <- M(s): a method M consumes information s and produces an explanation of information i with utility u for an inspector. Stephen Wormald led the work at the University of Florida. I am the second author of five.
Our paper in Artificial Intelligence Review, a Springer journal, introduces the Qi-Framework, a mathematical syntax for describing what an explanation of a machine learning model is. Stephen Wormald led the work at the Florida Institute for National Security under Dr. Damon Woodard. I am the second author, alongside Kristian O'Connor and Dr. Olivia Dizon-Paradis.
The name is short for Question-Information. An inspector, anyone from a CEO to an ML engineer, asks a question about a model. The framework assumes that question maps to an explanation that helps the inspector predict some information i with less error and less uncertainty. The syntax writes this as E{i,u} <- M(s): a method M consumes a set of information s and produces an explanation of information i that carries utility u for its inspector. Four variables. That compactness is the point.
Interactive demo: decompose an XAI method into the Qi-Framework
Every explanation can be written as E{i,u} <- M(s). Pick a method and watch each variable fill in. Hover a filled slot for its definition.
Why XAI needs a common syntax
XAI research produces hundreds of explanation methods yet offers few tools for comparing them quantitatively. Surveys and taxonomies sort methods into static categories such as post-hoc or model-agnostic, which describe the field without helping anyone measure whether an explainability need is actually met. The paper proposes syntax as the missing layer.
Policy makers, healthcare providers, and defense contractors need to understand model behaviors, and XAI has answered with saliency maps, surrogate models, prototype networks, and partial dependence plots. The trouble is comparison. Surveys and taxonomies sort methods into categories like post-hoc or model-agnostic, which helps a reader browse the field and does little to show whether a specific explainability need is met. Burkart and Huber named the gap directly: the field lacks a standard procedure to measure, quantify, and compare explainability.
Static labels also age badly. Loss curves explain the training process itself, so the post-hoc and inherent labels both sit awkwardly on them. Programming languages solved a similar problem with syntax, defining general variable classes that extend to programs nobody has written yet. The Qi-Framework borrows that move. Shared syntax turns a pile of methods into a field that can measure itself.
Decomposing explanations into sub-components
The framework was abstracted empirically from 184 explanations found across 66 XAI publications. Each explanation was decomposed into the spatial information it touches (input, latent, and output spaces), the function components involved (from single weights to whole architectures), and derived metrics such as attribution scores.
The syntax was abstracted from evidence. The team reviewed 66 XAI publications, found through highly cited surveys and popular GitHub repositories, and extracted 184 individual explanations from them. That count includes explicate explanations the authors named directly, plus secondary ones latent in a method or its figures. Literature review under Stephen was my own first job when I joined the lab in fall 2023, and this paper shows where patient reading can lead.
Each explanation was then decomposed with the same questions. What input types does it use? What function components are responsible for it? What subsidiary information does it need? The answers cluster into three families: spatial information (instances, perturbations, clusters, and full sets drawn from a model's input, latent, and output spaces), function components on five scales from single weights up to whole architectures, and derived metrics such as additive feature attributions, prototype similarity, and TCAV scores.
How to compare XAI methods with the framework
Explanation types in the Qi-Framework are compared through the information they explain and the utility they provide, which the review sorts into eight categories including importance, justification, and surrogacy. Writing methods like LIME in one shared notation exposes overlaps, and it exposes gaps that no existing method fills.
Explanations of the same information can serve inspectors differently, so the framework treats utility as a first-class variable. The review found eight recurring utility types: representation, justification, importance, logical clarity, reduction, surrogacy, examples, and counterfactuals. Of the 184 explanations, 57 percent explained spatial information, 25 percent explained function components, and 18 percent derived secondary metrics.
The payoff: any method can be written in the same notation and compared cell by cell. The paper works through LIME: from an input, a perturbed input set, and the perturbed outputs, LIME produces a linear surrogate model, and an association step maps the surrogate's weights onto input dimensions to explain their importance. Once everything shares a grid, gaps become visible. None of the 184 explanations ranked the importance of decision traces. The discussion then defines local, spatial, transversal, and global explanation effects in terms of changes in an inspector's error and uncertainty, sketching a route to quantitative XAI benchmarks.
Interactive demo: the whitespace analysis, information vs utility
Writing all 184 reviewed explanations in one syntax turns the field into a grid: rows are the information being explained, columns are the utility the explanation provides. Click a filled cell for an example method. Hover an empty cell to find open research territory.
Cells are illustrative groupings of the paper's finer-grained analysis; Appendix B of the paper expresses all 184 explanations in the full syntax.
Who this framework is for
Researchers, practitioners, and auditors in high-stakes fields are the intended users. The syntax gives practitioners a vocabulary for stating explainability needs, gives researchers a way to describe and generate methods, and gives the broader field a path toward the quantitative benchmarks that XAI has lacked.
Practitioners come first. Someone deploying a model in healthcare, finance, or national security needs to state an explainability requirement precisely, then check whether a candidate method meets it. The framework's vocabulary of information sets and utilities is built for exactly that. Researchers get a complementary use: describing new methods in a form others can compare against, and finding unexplained territory worth a paper. The future directions point toward explanation databases, shared XAI dashboards, and explainability thresholds a model must clear before deployment.
The framework is also a map I still use. My research at FINS sits on its inherently interpretable side, working with differentiable logic gate networks, models whose nodes learn logic gates a human can follow. eXpLogic, a later paper from our lab that generates saliency maps for those networks, walks ground this framework surveyed first. Second author was the right seat for this one. I got to watch a field being handed a grammar.
Abstract
Policy makers, healthcare providers, and defense contractors need to understand many types of machine learning model behaviors. While eXplainable Artificial Intelligence (XAI) provides tools for interpreting these behaviors, few frameworks, surveys, and taxonomies produce succinct yet general notation to help researchers and practitioners describe their explainability needs and quantify whether these needs are met. Such quantified comparisons could help individuals rank XAI methods by their relevance to use-cases, select explanations best suited for individual users, and evaluate what explanations are most useful for describing model behaviors. This paper collects, decomposes, and abstracts subcomponents of common XAI methods to identify a mathematically grounded syntax that applies generally to describing modern and future explanation types while remaining useful for discovering novel XAI methods. The resulting syntax, introduced as the Qi-Framework, generally defines explanation types in terms of the information being explained, their utility to inspectors, and the methods and information used to produce explanations. Just as programming languages define syntax to structure, simplify, and standardize software development, so too the Qi-Framework acts as a common language to help researchers and practitioners select, compare, and discover XAI methods. Derivative works may extend and implement the Qi-Framework to develop a more rigorous science for interpretable machine learning and inspire collaborative competition arcoss XAI research.
Frequently Asked Questions
How do I choose an XAI method?
Start from the question you need answered and the information it concerns. If you need to know which inputs drove a prediction, input saliency methods like LIME, SHAP, or GradCAM apply. If you need to understand what a model learned internally, state activation methods that interpret the latent space fit better. Our review found input importance and function representation are the most commonly served needs; many other needs have few or no methods.
What is a taxonomy of explanations in machine learning?
A taxonomy sorts explanation methods into categories, typically by labels like model-agnostic versus model-specific or local versus global. The paper compares over a dozen prior surveys, taxonomies, and method papers and finds most sort methods categorically, apply mainly to historical methods, and avoid mathematical notation. The Qi-Framework keeps the descriptive goal while replacing fixed categories with variables researchers can instantiate for methods that do not exist yet.
What is the difference between post-hoc and inherently interpretable models?
Post-hoc methods explain a model after training, while inherently interpretable models are built so their structure can be read directly. The review splits post-hoc work into input saliency and state activation methods, and inherent explainability into structural approaches (prototype layers, semantic bottlenecks, logic-based decisions) and training-based approaches that constrain the loss function so the latent space becomes interpretable.
Where can I read the Qi-Framework paper?
The version of record appeared in Artificial Intelligence Review in May 2025 under DOI 10.1007/s10462-025-11216-8, and an earlier preprint is on Research Square under a CC BY 4.0 license. Appendix B of the paper expresses all 184 reviewed explanations in the Qi-Framework syntax, which is the fastest way to see how the notation handles a method you already know.
Can the Qi-Framework generate new XAI methods?
Yes, and that is one of its novelty claims. Tabulating existing methods in the shared syntax works like a whitespace analysis: empty cells mark information no current method explains. The paper points to concrete examples, such as ranking the importance of decision traces in classification, and outlines future work on explanation databases, shared XAI dashboards, and deployment thresholds built on the framework.
BibTeX
@article{wormald2024xaisyntax,
title={Abstracting General Syntax for XAI after Decomposing Explanation Sub-Components},
author={Wormald, Stephen and Maldaner, Matheus Kunzler and O'Connor, Kristian and Dizon-Paradis, Olivia P. and Woodard, Damon L.},
journal={Artificial Intelligence Review},
publisher={Springer},
year={2024},
doi={10.1007/s10462-025-11216-8}
}