GLARE: A Natural Language Interface for Querying Global Explanations

Researchers introduce GLARE, an LLM-based interactive system that translates natural language questions into SQL queries to make global explanations from AI vision models more accessible and usable. The system bridges the gap between complex, static explanation artifacts and human-centered interpretability by enabling users to ask targeted questions about model behavior without needing technical expertise.

GLARE addresses a fundamental challenge in explainable AI (XAI): while global explanations help researchers understand vision model behavior across datasets and classes, their complexity and static presentation limit practical adoption. Traditional global explanations require users to navigate dense technical outputs, creating friction between what explainability systems can provide and what practitioners actually need.

This work emerges from a broader trend toward human-centered AI development where interpretability serves real user needs rather than theoretical understanding. As computer vision models become increasingly deployed in high-stakes domains—healthcare, autonomous systems, content moderation—the ability to interrogate model decisions dynamically has become essential. The shift from static artifacts to interactive query systems reflects growing recognition that one-size-fits-all explanations fail to address diverse stakeholder questions.

The practical impact extends across multiple constituencies. Data scientists and ML engineers gain faster insight into model failures and biases without maintaining separate explanation infrastructure. Organizations deploying vision systems can more efficiently audit behavior for regulatory compliance. The SQL-mediated approach proves particularly valuable because it separates logical query specification from presentation, allowing different stakeholders—auditors, domain experts, non-technical decision-makers—to interact with identical underlying data through their preferred interface.

Looking ahead, the critical evaluation metrics around intent interpretation and robustness to linguistic errors suggest maturity approaching production deployment. Success here could catalyze broader adoption of LLM-mediated query systems across other domains requiring explanation, from NLP models to recommendation systems, establishing new standards for XAI usability.

• →GLARE translates natural language questions into SQL queries, democratizing access to global explanations without requiring technical expertise.
• →LLM-mediated querying enables flexible, targeted analysis rather than static explanation artifacts, improving practical usability.
• →The system supports diverse stakeholder needs—auditors, domain experts, and non-technical users—through unified underlying data.
• →Evaluation covers intent interpretation, generalization, and linguistic robustness, indicating production-readiness potential.
• →This work exemplifies a broader industry shift toward human-centered explainability in high-stakes AI deployment.