Towards Symbolic XAI -- Explanation Through Human Understandable Logical Relationships Between Features

Explanations can go beyond highlighting individual features: Traditional XAI methods show which input features matter most (like highlighting words in text or pixels in images), but this paper demonstrates that AI models often use abstract reasoning patterns—logical relationships between features—that better match how humans actually solve problems. This suggests current explanations may be incomplete.

The framework works across different AI applications: The researchers tested Symbolic XAI on three distinct domains (natural language processing, computer vision, and quantum chemistry), showing the approach is generalizable rather than limited to one type of AI system, which strengthens its potential practical value.

Explanations can be customized and expressed as human-readable logic: Unlike black-box explanations, Symbolic XAI represents model reasoning as logical formulas that users can understand and customize based on their needs—making it potentially more useful for domain experts who need to verify whether an AI system's reasoning is sound for their specific application.

This addresses a gap between how AI explains itself and how humans think: The paper identifies that most XAI methods explain decisions at only one level of abstraction, whereas humans typically reason through multiple layers of logic. By capturing these intermediate reasoning steps, Symbolic XAI could build greater trust by showing work that feels more familiar to human decision-making.