Source-linked AI summary
LumiXAI: A Modular Full-Stack Framework for Feature Attribution
Alfio Ferrara, Lorenzo Gatta, Sergio Picascia, Elisabetta Rocchetti
TL;DR
Attribution software is fragmented across modalities, interfaces, and extensibility, limiting unified exploration. LumiXAI addresses this with a modular full-stack framework that combines attribution methods, bidirectional GUI exploration, plug-ins, and three access tiers. The paper presents this as a reproducible system for inspecting, repeating, and extending attribution workflows, while noting limits in current coverage and in attribution methods themselves.
Problem
Existing attribution tools specialize along narrow axes and often require code or method familiarity, leaving interfaces, modalities, and workflows fragmented.
Method
LumiXAI combines classification and generative attribution, bidirectional GUI exploration, registered model and method plug-ins, and GUI, SDK, and extender access through one backend.
Results
LumiXAI operationalizes established attribution techniques across multiple model families and access modes in workflows that can be inspected, repeated, and extended.
Takeaways & Limitations
The framework consolidates attribution analysis for non-programmers, developers, and extenders within a unified system.
Takeaways & Limitations
Current implementation covers only a subset of tasks, models, and attribution methods, while attribution scores remain diagnostic signals rather than causal explanations.
Abstract
from arXiv · showhide
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.
1 Introduction
LumiXAI addresses fragmented attribution tooling with a modular framework spanning multiple model families, interaction styles, and user access modes. It combines bidirectional exploration, extensibility, and persistent analyses in one workflow.
- Attribution tooling often specializes in narrow axes, leaving researchers to reconcile incompatible interfaces, output formats, and deployment assumptions.
- LumiXAI supports feature attribution across text and image classification, autoregressive text generation, and text-to-image diffusion.
- Registered model and attribution plug-ins, independent Docker services, and persisted metadata and outputs make analyses extensible, inspectable, and reproducible.
- The framework serves web GUI, Python SDK, and smart-batch users through the same backend.
- Text-to-image exploration supports both token-to-image heatmaps and reverse image-to-token inspection.
2 Related Work
Existing attribution tools provide varied method and modality coverage but commonly offer limited interactive support, extensibility, or generative scope. LumiXAI combines generative attribution across text and text-to-image diffusion with a GUI, extension interfaces, and shared access tiers.
- Existing tools tend to specialize across a small number of dimensions, including modality, generative attribution, GUI support, extensibility, and audience.
- General-purpose libraries offer broad attribution backends but limited interactive support, while many tools remain classification-oriented or code-based.
- Inseq and ICX360 support generative text attribution but are text-only and API-driven.
- LumiXAI combines generative attribution for text and text-to-image diffusion with an interactive GUI, documented extension interfaces, and shared access tiers.
3 The LumiXAI Framework
LumiXAI separates model access, attribution computation, clients, and deployment through a REST-based plug-in architecture. Its supported interfaces and persistent storage support GUI, programmatic, batch, and extension workflows, while framework overhead remains unevaluated.
- Architecture Overview: A REST API isolates model-specific and compute-intensive functionality, allowing the web interface, Python SDK, and batch workflows to share one backend.
- Extensibility: Wrappers and attributors decouple model access from explanation computation and let new components be added through subclasses, applicability declarations, and registry entries.
- Architecture Overview: The system deploys independent backend, frontend, and documentation services, with model weights loaded only by the backend.
- Storage: SQLite stores structured job metadata, while JSON files store larger attribution payloads for later inspection, export, or transfer.
- Extensibility: Modality-agnostic contracts represent tokens, patches, or superpixels as input features and return standardized attribution objects.
- Supported Models and Access: The release includes ten established attribution methods and access modes for GUI use, Python SDK workflows, smart batching, and framework extension.
- Deployment: Runtime is mainly determined by the selected model and attribution method; dedicated measurement of framework overhead is left to future work.
4 Case Studies
LumiXAI demonstrates attribution workflows for text classification and text-to-image diffusion, combining token- or pixel-level inspection with bidirectional exploration. The case studies show how users can examine which inputs drive predictions and how prompt tokens influence generated image regions.
- Text Classification: Integrated Gradients colors Civil Comments tokens by signed attribution toward the toxicity classifier’s predicted class.The case study uses unitary/toxic-bert and examines whether predictions rely on identity terms, offensive language, or quoted stereotypes.
- Text Classification: In the example insult, “losers” dominates attribution to the predicted INSULT class at 51%.The visualization colors tokens according to their attribution scores.
- Text-to-Image Diffusion: DAAM attributes generated image pixels to prompt tokens during text-to-image diffusion.The case study generates images from prompts pairing a subject with an artistic style.
- Text-to-Image Diffusion: Selecting a prompt token overlays its token-to-image heatmap, while selecting an image region reveals the responsible prompt tokens.This bidirectional interaction supports comparison of content and style influence and inspection of their entanglement.
5 User Evaluation
The user evaluation examined whether LumiXAI supports usable attribution inspection for experts and non-experts across four task types. Participants rated the system positively, while feedback identified onboarding, heatmap scaling, and compatibility information as accessibility priorities.
- Evaluation Design: The evaluation targeted usability and workflow effectiveness rather than the intrinsic quality of explanations.Explanation quality was treated as dependent on the external attribution libraries used.
- Setup: The study recruited 12 participants: 3 with machine-learning or NLP experience and 9 non-experts.Expertise was determined using background questions about training or evaluating models and using attribution or explainability methods.
- Setup: Participants completed tutorials, a comprehension check, and 8 tasks spanning text classification, image classification, autoregressive text generation, and text-to-image diffusion.Each task used a fixed model and attribution method with a provided input.
- Results: Participants judged LumiXAI positively, especially for switching between models and attribution methods and understanding model behaviour.Ratings were consistently higher among expert users.
- Results: Clearer onboarding, transparent heatmap colour scaling, and explicit method–model compatibility were identified as improvements for non-expert accessibility.These priorities came from qualitative participant feedback.
6 Conclusion
The paper concludes that LumiXAI provides reproducible, interactive feature-attribution analysis across multiple model families, access modes, and user profiles. Its plug-in and service-based design supports inspection, repetition, extension, debugging, and exploratory interpretation.
- Conclusion: LumiXAI operationalizes established attribution techniques in a unified system spanning multiple model families, access modes, and user profiles.The framework combines a plug-in architecture, service-based deployment, persistent storage, and complementary interfaces.
- Conclusion: The framework turns ad hoc attribution scripting into analyses that can be inspected, repeated, and extended.Case studies illustrate targeted debugging and exploratory interpretation across classification, generation, and text-to-image diffusion.
Limitations
LumiXAI’s current implementation covers only a subset of possible tasks, models, and attribution methods. Its attribution outputs also inherit methodological limitations, while multiuser robustness and broader evaluation remain unresolved.
- Scope: Current coverage includes only a subset of possible tasks and models, with vision-language architectures requiring additional wrappers and visualization components.Supported attribution methods are representative rather than exhaustive.
- Attribution Methods: Attribution scores vary with baselines, tokenization, perturbation units, aggregation choices, and preprocessing.They should therefore be treated as diagnostic signals rather than causal explanations, especially for generative models.
- System Robustness: The system was developed and tested primarily under single-user conditions, and concurrent multiuser loading caused bugs during evaluation.Robustness under concurrent multiuser load remains future work.
- Evaluation Scope: The user evaluation tested guided task completion and visualization interpretation but not long-term open-ended research use or dedicated systems performance.The study also did not isolate framework overhead from model inference and attribution time.
Ethics Statement
The evaluation protected participants’ autonomy and privacy through informed consent, voluntary participation, withdrawal rights, and anonymised aggregate reporting.
- Written informed consent was obtained before participants began the tutorial phase.
- Participants could withdraw from the study at any point without giving a reason.
- Questionnaire responses and free-text feedback were anonymised at collection and reported only in aggregate.
A Additional Case Studies
LumiXAI demonstrates attribution analysis for image classification and open-ended text generation through visual, interactive case studies. The examples show spatial attribution for digit recognition and token-level attribution for generated text.
- Image Classification: A Vision Transformer trained on MNIST predicts digit 3, while Grad-CAM overlays attribution on the visual input.The strongest activations align with the upper, middle, and lower curves defining the handwritten digit.
- Image Classification: The image view enables inspection of whether the classifier relies on digit strokes rather than background artifacts.
- Open-Ended Text Generation: LumiXAI uses per-step Integrated Gradients to trace how earlier prompt tokens contribute to autoregressive generation.The case study uses deterministic GPT-2 continuations from Wikipedia-derived BOLD prompt prefixes.
- Open-Ended Text Generation: Users can select any generated token and inspect which earlier tokens contributed most to it.
- Open-Ended Text Generation: For the generated token Islamic, Queen contributes negatively (−50%), whereas Lat contributes positively (+73%) in the signed attribution trace.