GLARE: A Natural Language Interface for Querying Global Explanations

Modern global explanations for deep vision models are often monolithic and overwhelming, making it difficult for users to extract specific insights. The authors address this usability gap by asking: Can we treat global explanations as a queryable database, allowing users to interrogate model behavior through natural language rather than manually parsing static artifacts?

GLARE (Global Language-based Analysis and Retrieval of Explanations) treats local, concept-based explanations as entries in a relational database. The system uses a fine-tuned Large Language Model (LLM) to act as a semantic parser. When a user asks a question, the LLM maps the intent to one of 24 specialized SQL templates. To ensure reliability, the authors use a technique called "SQL-fence loss masking," which forces the model to learn the relational algebra of explanation querying while ignoring irrelevant text. This approach allows the system to return precise statistics, supporting evidence images, and visualizations based on the user's specific query.

GLARE demonstrates high accuracy and robustness across multiple evaluation axes. In tests on the ADE20K dataset, the system achieved over 95% accuracy in matching user intent to correct SQL queries. Notably, the model exhibits strong zero-shot transfer capabilities; when tested on the Pascal VOC dataset—which contains an entirely different object vocabulary—the system maintained high performance without any retraining. Furthermore, the model proved significantly more robust to linguistic noise, such as spelling errors and synonym variations, compared to traditional regex-based baseline methods.

This work shifts the paradigm of Explainable AI (XAI) from viewing explanations as static, "take-it-or-leave-it" documents to treating them as interactive, queryable data. By lowering the barrier to entry for non-experts, GLARE enables more efficient auditing of model biases and failure modes, facilitating a more human-centered approach to model transparency.