Addressing Explainability of Generative AI using SMILE (Statistical Model-agnostic Interpretability with Local Explanations)
Zeinab Dehghani
TL;DR
This work presents gSMILE, a model-agnostic, perturbation-based framework that extends the SMILE paradigm to generative AI, unifying explainability for both large language models and instruction-based image editing. By perturbing input prompts, measuring output distribution shifts with Wasserstein-based distances, and fitting weighted local surrogates, gSMILE yields token-level attributions and intuitive heatmaps that reveal how prompts drive generation and editing decisions. The framework is evaluated with rigorous metrics—accuracy, stability, consistency, and fidelity—across multiple LLMs and diffusion-based editors, demonstrating robust, human-aligned explanations and generalisability to state-of-the-art architectures. gSMILE advances transparent, reliable deployment of multimodal generative AI by offering principled, model-agnostic explanations without requiring access to internal model parameters, albeit with higher computational cost. The findings highlight the method's potential for improving prompt design, bias detection, and accountability in high-stakes applications, and point to future work in richer perturbations, causal verification, and broader modality support.
Abstract
The rapid advancement of generative artificial intelligence has enabled models capable of producing complex textual and visual outputs; however, their decision-making processes remain largely opaque, limiting trust and accountability in high-stakes applications. This thesis introduces gSMILE, a unified framework for the explainability of generative models, extending the Statistical Model-agnostic Interpretability with Local Explanations (SMILE) method to generative settings. gSMILE employs controlled perturbations of textual input, Wasserstein distance metrics, and weighted surrogate modelling to quantify and visualise how specific components of a prompt or instruction influence model outputs. Applied to Large Language Models (LLMs), gSMILE provides fine-grained token-level attribution and generates intuitive heatmaps that highlight influential tokens and reasoning pathways. In instruction-based image editing models, the exact text-perturbation mechanism is employed, allowing for the analysis of how modifications to an editing instruction impact the resulting image. Combined with a scenario-based evaluation strategy grounded in the Operational Design Domain (ODD) framework, gSMILE allows systematic assessment of model behaviour across diverse semantic and environmental conditions. To evaluate explanation quality, we define rigorous attribution metrics, including stability, fidelity, accuracy, consistency, and faithfulness, and apply them across multiple generative architectures. Extensive experiments demonstrate that gSMILE produces robust, human-aligned attributions and generalises effectively across state-of-the-art generative models. These findings highlight the potential of gSMILE to advance transparent, reliable, and responsible deployment of generative AI technologies.