U-CECE: Multi-Resolution Framework for AI Explainability

Researchers propose U-CECE, a unified framework for conceptual counterfactual explanations in AI, addressing the trade-off between expressivity and efficiency. The model-agnostic system operates at three levels: atomic concepts, relational sets-of-sets, and structural graphs. At the structural level, it supports both transductive mode via supervised Graph Neural Networks (GNNs) and inductive mode via unsupervised graph autoencoders (GAEs). The framework adapts to data regimes and compute budgets, offering a scalable solution for explainability. The paper is available on arXiv (2604.08295).