LumiXAI: A Modular Full-Stack Framework for Feature Attribution

Alfio Ferrara; Lorenzo Gatta; Sergio Picascia; Elisabetta Rocchetti

cs.SEAug 25, 2026

LumiXAI: A Modular Full-Stack Framework for Feature Attribution

Authors: Alfio Ferrara, Lorenzo Gatta, Sergio Picascia, Elisabetta Rocchetti

Organizations: Department of Literary Studies, Philology and Linguistics, Università degli Studi di Milano, Milan, Italy · Department of Computer Science, Università degli Studi di Milano, Milan, Italy

Abstract

Feature attribution is a central tool of model interpretability, yet the software through which it is applied remains fragmented: individual tools specialize along narrow axes, such as a single modality, a code API or a GUI, or a fixed rather than extensible method set, and rarely combine these strengths. Moreover, many explainability tools are designed primarily for domain experts, requiring programming skills or familiarity with attribution methods that can make them difficult for non-expert users to access. In this article, we present LumiXAI, a modular full-stack framework that consolidates attribution analysis into a single system. It couples classification and generative attribution with an interactive GUI supporting bidirectional exploration, a plug-in architecture for registering new models and methods, and three access tiers serving non-programmers, developers, and extenders from one backend. Its contribution is a system that operationalises established attribution methods under one interface, one interaction model, and one persistence layer, with containerised services and persistent results making analyses reproducible across machines.

Figures & tables

Figure 1: Bidirectional exploration of text-to-image attribution in LumiXAI. Top (image $\rightarrow$ token): hovering a pixel reads that point from every prompt token’s attention matrix and recolors the prompt cards. Bottom (token $\rightarrow$ image): clicking a prompt token overlays its attribution heatmap on the generated image.

Tool	Scope	Gen. Attr.	GUI	Ext.	Audience
Captum	TC, TG, IC
SHAP / LIME	TC
OmniXAI	TC, IC
AIX360	TC, IC
AllenNLP Interpret	TC, TG
LIT	TC, TG

Table 1: Comparison of attribution tools. Scope : TC = text classification, TG = text generation, IC = image classification, T2I = text-to-image generation. Generative attribution : first-class attribution over the generative decoding process (autoregressive or diffusion); limited to masked-token or single-step prediction, or salience confined to particular seq2seq settings; classification or regression outputs only. GUI : whether it ships an interactive visual interface, e.g. standalone application or interactive notebook widget. Ext. : if it is designed for adding new attribution methods or model types. Audience : extender (adds methods or models), developer (applies methods through a code API), non-programmer (works through the GUI alone). Cells reflect released versions as of July 2026 (Captum 0.9.0, SHAP 0.52.0, Inseq 0.7.1, LIT 1.3.1, CafGa 0.0.6, ICX360 0.1.0).

Figure 2: LumiXAI architecture. The backend exposes a REST API, orchestrates wrapper and attributor plug-ins, retrieves models from the Hugging Face Hub, and stores metadata in SQLite with attribution payloads as JSON files. Backend, frontend, and documentation are deployed as independent containers, and both the web GUI and client scripts consume the same API.

Figure 3: Text classification in LumiXAI. For the comment “ha ha you guys are a bunch of losers.” , classified as insult by unitary/toxic-bert , tokens are colored by Integrated-Gradients attribution toward the predicted class; losers dominates the attribution (51%).

Figure 4: Text-to-image attribution in LumiXAI. For the prompt “a impressionism painting of a pizza” , generated by Stable Diffusion XL , selecting pizza overlays its DAAM heatmap (token $\rightarrow$ image), while token percentages report attribution at the selected pixel (image $\rightarrow$ token).

Question	Overall	Experts	Non-experts
Ease of use	4.92	5.67	4.67
Learnability	5.17	6.33	4.78
Enjoyability	4.75	5.67	4.44
Integration	5.33	5.67	5.22
Helpfulness	5.25	6.00	5.00
Reuse intention	4.25	4.33	4.22

Table 2: Mean user-study ratings (7-point Likert scale) per questionnaire item, overall and by expertise (

n=12

; 3 experts, 9 non-experts).

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 5: Image classification in LumiXAI. A ViT model trained on MNIST predicts the digit 3 ; the Grad-CAM overlay highlights the image regions that most support the predicted class, with attribution concentrated on the main strokes of the digit.

Figure 6: Open-ended generation in LumiXAI. Continuing the BOLD prefix “Queen Latifah is a contralto, and she” with gpt2 , the selected token Islamic is explained by signed Integrated-Gradients scores over the prior context; Queen contributes negatively ( $-50\%$ ), while Lat contributes positively ( $+73\%$ ).

Explainable Artificial Intelligence

Explainable artificial intelligence develops ways to make AI decisions understandable, testable, and useful to people.

See all 326 papers

Hypothesis Class Determines Explanation: Why Accurate Models Disagree on Feature Attribution
Mar 16, 2026Thackshanaramana BFeature Attribution Explainability Evaluation
SRM Institute of Science and Technology, India
Measuring Explainer Stability via Attribution Separability
Aug 3, 2026Eddie Conti, Álvaro Parafita, Axel BrandoFeature Attribution Explanation Stability
Barcelona Supercomputing Center, Barcelona, Spain
ProToMEx: Rapid, Interpretable Explanations via Structured Representations
Sep 2, 2026Athina Georgara, Adarsh Valoor, Sarvapali D. RamchurnExplainable Artificial Intelligence Interpretable ML
University of Southampton, Hampshire, England, UK

Mar 16, 2026J/K move · Enter open · S save

Hypothesis Class Determines Explanation: Why Accurate Models Disagree on Feature Attribution

Thackshanaramana B

SRM Institute of Science and Technology, India

Feature Attribution Explainability Evaluation Model Selection

Open paper AskPDF arXiv

The assumption that prediction-equivalent models produce equivalent explanations underlies many practices in explainable AI, including model selection, auditing, and regulatory evaluation. In this work, we show that this assumption does not hold. Through a large-scale empirical study across 24 datasets and multiple model classes, we find that models with identical predictive behavior can produce substantially different feature attributions. This disagreement is highly structured: models within the same hypothesis class exhibit strong agreement, while cross-class pairs (e.g., tree-based vs. linear) trained on identical data splits show substantially reduced agreement, consistently near or below the lottery threshold. We identify hypothesis class as the structural driver of this phenomenon, which we term the Explanation Lottery. We theoretically show that the resulting Agreement Gap persists under interaction structure in the data-generating process. This structural finding motivates a post-hoc diagnostic, the Explanation Reliability Score R(x), which predicts when explanations are stable across architectures without additional training. Our results demonstrate that model selection is not explanation-neutral: the hypothesis class chosen for deployment can determine which features are attributed responsibility for a decision.

Abstract

Explainable Artificial Intelligence

Explore similar work

Hypothesis Class Determines Explanation: Why Accurate Models Disagree on Feature Attribution

Measuring Explainer Stability via Attribution Separability

ProToMEx: Rapid, Interpretable Explanations via Structured Representations