GLARE: A Natural Language Interface for Querying Global Explanations

Researchers have introduced GLARE, a natural language interface that lets users query global explanations of black-box image classifiers without wrestling with low-level data structures. The system, described in a paper submitted 18 June 2026, uses a large language model to translate plain-English questions into structured SQL queries over local explanation data. [1][2] Global explanations are essential for understanding how vision models behave across datasets, classes, and decision contexts, but their complexity often blocks practical exploration. Users typically want targeted answers to specific questions rather than static artifacts, the authors note. GLARE addresses this by positioning an LLM as a mediator: a user asks a question in natural language, the LLM converts it into a SQL query, and the system returns statistics-augmented natural language responses alongside intent-aligned visualizations. [1][2] The interface also surfaces supporting local explanations for each query, giving users a layered view of model behavior. The paper evaluates GLARE on four dimensions: intent interpretation, query mapping accuracy, generalization to novel queries and datasets, and robustness to linguistic errors. The results indicate that LLM-mediated querying substantially improves the accessibility and usability of global explanations for human-centered explainable AI. [1][2] The work arrives as the broader AI community continues to grapple with transparency in black-box systems. While GLARE focuses on image classifiers, the challenge of making model internals interpretable spans domains. In molecular biology, for instance, roughly 1,600 transcription factors in the human genome regulate gene expression by binding to specific DNA sequences, and understanding their combinatorial logic remains a parallel interpretability problem. [7] Similarly, efforts to consolidate heterogeneous datasets for catalyst informatics have shown that transfer learning between datasets can improve model performance, but only when researchers can interrogate what a model has learned across training sources. [4] The GLARE system is hosted on arXivLabs, a framework that lets collaborators develop and share new arXiv features directly on the platform. Both individuals and organizations working with arXivLabs have accepted the site's values of openness, community, excellence, and user data privacy. The paper's code and media links also point to Hugging Face, a repository widely used for sharing machine-learning models and datasets. [1]