GFGE: Unifying Explainable AI Methods through an Interpretation Framework

Jinfeng Zhong

cs.AIOct 4, 2026

GFGE: Unifying Explainable AI Methods through an Interpretation Framework

Authors: Jinfeng Zhong

Organizations: LIASD, Université Paris 8, 93200 Saint-Denis, France

Abstract

Explainable artificial intelligence (XAI) encompasses methods that draw on different sources of information and address different explanatory needs. A common framework is needed to describe how this information becomes evidence and is communicated as an explanation for a particular recipient. We propose the General Framework for Generating Explanations (GFGE), grounded in interpretative frameworks and the complementary activities of \emph{sense-reading} and \emph{sense-giving}. Its conceptual foundation is the Interpret/Explain Schema (IES), which connects an analyst's interpretation of system evidence, the communication of a selected account, and the recipient's interpretation of that account. GFGE operationalises this schema through five roles: data interpretation, model interpretation, output interpretation, optional post-hoc analysis, and aggregation. A role-typed operation graph records method-specific dependencies, while evidence records retain the sources, assumptions, and limitations of explanatory claims. The explanatory question, audience, and context guide the procedure. We instantiate GFGE for attribution, surrogate, counterfactual, concept and prototype, intrinsic rule, argumentation, and language-model methods. These instantiations show how intrinsic, post-hoc, and hybrid workflows can be represented through the same roles while preserving their distinct evidential requirements. GFGE provides a common basis for analysing explanation workflows, tracing communicated claims to their evidence, and identifying unresolved explanatory dependencies.

Figures & tables

Panel of Figure 1 — Figure 1: An analyst reads evidence through an interpretative framework, gives an account to a recipient, and the recipient reads that account through a different framework. This is our conceptual illustration of sense-reading and sense-giving [ 20 , 5 ] .

Review	Primary organising question	GFGE’s distinct analytical question
Schwalbe and Finzel [ 58 ]	Which properties distinguish methods and guide their selection?	Which interpretative operations and information flows generate a particular explanation?
Wang et al. [ 72 ]	Which stakeholders need which explanations at which lifecycle stage?	How does a selected method transform data, model, and output evidence into an account for that stakeholder?
Cousineau and Dara [ 17 ] ; Elsayed et al. [ 23 ]	How should techniques be classified or selected by mechanism and interpretability level?	How are intrinsic, post-hoc, and hybrid techniques instantiated in the same modular procedure?
Mangold et al. [ 43 ]	How should XAI systems be designed and evaluated with human users?	Which generation steps precede that evaluation, and where does recipient interpretation enter?
Cai et al. [ 12 ]	How can LLM interpretability support intervention and model improvement?	How do internal evidence, intervention tests, and communicated explanations fit into a framework spanning model families?

Table 1: Scope of related XAI reviews and the complementary role of GFGE.

Panel of Figure 2 — Figure 2: Interpretation and explanation in an AI system. The analyst reads information from data, model, and output; an interface conveys an account to the recipient, who interprets it through their own mental model. Solid arrows denote computation or communication; dashed arrows denote information access.

Role	Input $\to$ output	Necessary assignment condition	Exclude and assign elsewhere
ID	$(\mathcal{D},z,c,q)\to E_{D}$	Extracts feature meaning, measurement, domain constraints, or feasible changes.	Desired outcome belongs to $q$ ; a model response belongs to IO.
IM	$(\mathcal{M},z,q)\to E_{M}$	Directly reads a mechanism, parameter, rule, leaf, prototype, or circuit actually used by the predictor.	A fitted surrogate or attribution computation belongs to PH, even if model internals accelerate it.
IO	$(y,\mathcal{M},z,q)\to E_{O}$	Records observed predictions, scores, or queried contrasts produced by $\mathcal{M}$ .	The requested alternative is in $q$ ; searching for inputs that attain it belongs to PH.
PH	$(\mathcal{M},z,y,q,E_{0})\to E_{H}$	Runs an additional explanatory analysis: perturbation, attribution, surrogate fitting, intervention test, or candidate search.	Simply reading an existing model rule is IM; wording an existing claim is A.
A	$(E,q,u)\to e$	Selects, orders, and renders supported claims with provenance for audience $u$ .	Any new model query, statistical estimate, or factual inference must be recorded in IO, PH, ID, or IM first.

Table 2: GFGE module contracts.

E_{0}=E_{D}\cup E_{M}\cup E_{O}

and

E=E_{0}\cup E_{H}

denote available records;

q

specifies the target.

Panel of Figure 3 — Figure 3: GFGE roles, evidence, and communication. (a) Computational operations supply evidence for analyst sense-reading; aggregation performs sense-giving, followed by recipient sense-reading. The role panel lists the vocabulary of $G_{f}$ , without imposing an execution order. Solid arrows denote evidence or communication; dashed arrows denote contextual inputs or feedback. (b) Selected dependency paths show PH before IM in CBM-zero and repeated PH operations in attribution-guided counterfactuals [ 71 , 76 ] . These paths omit other roles.

Family	Representative methods	GFGE roles and evidential issue
Feature attribution	SHAP, TreeSHAP, Integrated Gradients, DeepLIFT, LRP, Grad-CAM [ 41 , 40 , 63 , 60 , 7 , 59 ]	ID grounds features; IO records the prediction; PH computes contributions; A presents them. IM supplies structure for model-specific variants.
Local and global surrogates	LIME, surrogate trees, soft decision trees [ 55 , 18 , 27 ]	IO gathers model responses; PH fits an approximation; A reports its rules or coefficients with their scope.
Counterfactual and recourse	Wachter CF, DiCE, FACE, EECE, BRACE [ 69 , 49 , 52 , 73 , 25 ]	ID records mutable, plausible changes; $q$ fixes the alternative; IO checks outputs; PH searches candidates; A communicates the contrast. EECE also uses IM for forest structure.
Attribution-driven hybrid	SHAP-guided CF, CF-SHAP, recommendation CF [ 54 , 1 , 76 ]	PH composes attribution and search; ID constrains changes; IO records candidate model outputs; A combines findings and assumptions.
Concept and prototype	TCAV, ProtoPNet, ProtoTree, CBM-zero, FACE concept extraction [ 34 , 15 , 50 , 71 , 10 ]	ID grounds concepts or prototypes; IM exposes intrinsic concept mechanisms; PH tests post-hoc influence; A presents selected evidence.
Intrinsic rules and lists	Falling Rule Lists, CORELS, decision sets [ 70 , 3 , 36 ]	IM exposes predictive and active rules; IO records the result; ID grounds rule terms; A adapts the account to the question.

Table 3: Representative XAI families within the common GFGE procedure.

Method	Method-trait taxonomy	Stakeholder roadmap	GFGE trace and unresolved claim
CBM-zero [ 71 ]	Concept-based, model-access-dependent representation; local concept claims can be distinguished from global model traits.	Can pair a concept-level “what/why” question with an audience and stage.	PH constructs an invertible concept mapping; IM reads the converted predictor’s concept-to-output mechanism; IO records prediction preservation; ID must ground concept meaning. Recipient-specific A is unspecified. The order PH $\to$ IM and the semantic gap are explicit.
ReCalX [ 19 ]	Perturbation-based, post-hoc analysis with reliability as a relevant method property.	Can identify a debugging need and method for a prediction question.	IO records scores under perturbation; PH checks calibration, adjusts the queried evidence, then recomputes an explanation. A downstream account and its audience are unspecified. An attribution without the calibration assumption remains a flagged claim.
FaithLM [ 16 ]	Language explanation with iterative intervention-based assessment.	Can connect a recipient’s question to a language explanation method.	IO records rationales produced by $\mathcal{M}$ ; PH intervenes to test faithfulness and inform revision; A renders supported, recorded claims. This is a method-internal loop, not observed recipient feedback. Recipient understanding remains unspecified.

Table 4: The same methods viewed through two related frameworks and GFGE. Entries describe what each representation foregrounds; “unspecified” means the cited method does not supply that GFGE operation in the present analysis.

Workflow	Usual label	GFGE diagnosis
LIME/SHAP on latent features	Local, post-hoc	PH can compute scores while ID lacks a user-understandable mapping; A should state that limitation.
Attribution-guided counterfactual	Attribution or counterfactual	PH composes two operations; ID constrains changes, $q$ specifies the target, IO checks it, and A joins the claims.
EECE or causal recourse	Counterfactual, post-hoc	IM may supply forest structure; ID records feasible changes; PH searches candidates; A reports assumptions.
Intrinsic rule or argumentation model	Inherently interpretable	IM exposes reasoning structure, but IO selects the relevant contrast and A addresses the recipient.
Concept/prototype hybrid	Concept or prototype based	ID grounds semantics, IM or PH identifies influence, and A communicates the supported finding.
Generated LLM rationale	Natural-language explanation	New propositions are recorded as candidates and require evidence in ID, IM, IO, or PH; A renders supported claims. PH can test faithfulness to model behaviour.

Table 5: Explanatory gaps left implicit by taxonomic labels.

Explore similar work

Unraveling the Real Working Mechanism and Inherent Flaws of GAE: A Method for Interpreting Transformer Processes from an Economic Perspective
Sep 7, 2026Yongjin Cui, Xiaohui FanExplainable Artificial Intelligence Xai
Zhejiang University
Measuring What Matters: A Unified Evaluation Framework for GNN Explainability
Jul 6, 2026Francesco Paolo Nerini, Mirko Zaffaroni, Paolo Baracco +2Explainable Artificial Intelligence Gnn-Based Detectors
Sapienza University of Rome, Rome, Italy · Intesa Sanpaolo AI Research, Turin, Italy · Anti Financial Crime Digital Hub, Turin, Italy +1
CoAX: Cognitive-Oriented Attribution eXplanation User Model of Human Understanding of AI Explanations
Apr 30, 2026Louth Bin Rawshan, Zhuoyu Wang, Brian Y. LimXai Cognitive Science
National University of Singapore, Singapore

Sep 7, 2026, cs.AIJ/K move · Enter open · S save

Unraveling the Real Working Mechanism and Inherent Flaws of GAE: A Method for Interpreting Transformer Processes from an Economic Perspective

Yongjin Cui, Xiaohui Fan

Zhejiang University

Explainable Artificial Intelligence Xai Transformer Architectures Economies Data-Driven

Open paper AskPDF arXiv

We observe a phenomenon that current algorithmic research in the field of explainable artificial intelligence primarily pursues better performance on several proxy metrics. On the one hand, these proxy metrics themselves are more or less flawed and cannot properly measure the quality of methods. On the other hand, metric-oriented research approaches often lead to the neglect of the rationality and interpretability of the methods themselves. Explainable artificial intelligence is abbreviated as XAI. The metric-driven research paradigm has resulted in a lack of interpretability of the relevant XAI methods themselves. Accordingly, there is a need for interpretability research on XAI methods, which can be playfully referred to as XXAI. This paper is one of our works on XXAI. This paper takes Generic Attention-model Explainability (GAE), a widely influential model interpretation method , or rather, XAI method that represents an important technical route, as the research object, and explores the real working mechanism and flaws of this method as well as the technical route it represents. Based on the conclusions of this study, it may be necessary to re-examine or verify GAE-related methods and their domain applications. We argue that GAE is an interpretation method that focuses on the attention process. After pointing out the working mechanism and flaws of GAE, we propose Cumulative Asset Holdings (CAH), a more reasonable Transformer interpretation method integrating both process-based and feature-based ideas from an economic zero-sum games perspective. In addition, it is worth noting that our method is applicable to models with special tokens, where existing methods may suffer from limitations. The model simplification research method and the analysis of additive operations adopted in this study may provide inspiration for other research works in XAI.

Abstract

Explore similar work

Unraveling the Real Working Mechanism and Inherent Flaws of GAE: A Method for Interpreting Transformer Processes from an Economic Perspective

Measuring What Matters: A Unified Evaluation Framework for GNN Explainability

CoAX: Cognitive-Oriented Attribution eXplanation User Model of Human Understanding of AI Explanations