Research¶
SMILE—Statistical Model-Agnostic Interpretability with Local Explanations—is a family of local explainability methods for analysing black-box artificial-intelligence systems. The foundational method explains machine-learning classifiers by fitting a local surrogate model whose perturbed samples are weighted using statistical distance measures. Later work adapts this principle to spatial, generative, language, retrieval-augmented, and concept-based systems.
Research map¶
| Research area | Method or study | Explained system | Publication status |
|---|---|---|---|
| Core method | SMILE | Machine-learning and deep-learning classifiers | Peer-reviewed journal article |
| Vision and spatial AI | Point-cloud SMILE | Point-cloud neural networks | arXiv preprint |
| Vision and generative AI | Image-editing SMILE | Instruction-based image-editing models | arXiv preprint |
| Language and generative AI | gSMILE | Large language models | arXiv preprint |
| Retrieval-augmented generation | KG-SMILE | Knowledge-graph and GraphRAG systems | arXiv preprint |
| Concept-based explainability | ConceptSMILE | Concept-based explainable-AI methods | arXiv preprint |
| Financial language analysis | Local perturbation explanations derived from gSMILE | LLM financial-sentiment reasoning | Peer-reviewed conference chapter |
| Consolidated academic study | Generative-AI SMILE thesis | LLMs and instruction-based image editing | MSc thesis |
How the research family is organised¶
Core method¶
The foundational SMILE paper defines the statistical-distance-based local explanation framework for black-box classifiers. This is the principal methodological citation for general references to SMILE.
Vision and spatial AI¶
The point-cloud extension explains influential groups of 3D points and evaluates explanation fidelity, stability, and robustness. The image-editing extension instead perturbs natural-language editing instructions and measures how words or phrases affect generated visual changes.
Language and generative AI¶
gSMILE explains large-language-model behaviour by perturbing prompt components, measuring changes in model outputs with statistical distances such as Wasserstein distance, and fitting a locally weighted surrogate model. A related financial-sentiment study applies local perturbation explanations to behavioural and robustness analysis.
Retrieval-augmented generation¶
KG-SMILE explains the contribution of retrieved entities, relations, graph paths, and contextual evidence within knowledge-graph retrieval-augmented generation workflows.
Concept-based explainability¶
ConceptSMILE audits whether human-understandable concepts provide explanations that are faithful, locally representative, stable, and consistent.
Research resources¶
- Browse the complete publication catalogue
- Choose the correct citation for your use case
- Follow the reproducibility guidance
Publication metadata
The publication types, identifiers, links, and BibTeX records in this section are maintained from the project bibliography. Preprints and online-first records may later receive updated journal, conference, volume, issue, or pagination metadata.