Technical Summary: GLARE – A Natural Language Interface for Querying Global Explanations

1. Problem Statement

Deep vision models, while successful in high-stakes domains, remain opaque. While Explainable AI (XAI) offers local explanations (e.g., saliency maps for single instances) and global explanations (e.g., aggregated rules or prototypes summarizing model behavior across a dataset), a critical usability gap exists. Current global explanations are often monolithic, static artifacts containing thousands of logical rules or prototypes. Presenting this "explanation dump" to users induces cognitive overload, obscuring insights rather than revealing them.

Users typically do not need a static summary; they engage in an iterative, query-driven process to test specific hypotheses (e.g., "Does the model rely on background snow to classify wolves?"). Existing tools force users to manually filter and aggregate local explanations to answer such questions, creating friction that limits the practical utility of global insights. There is a need for an interface that allows users to interrogate global explanations using natural language without exposing them to low-level representations or requiring programming expertise.

2. Methodology

The authors propose GLARE (Global Language-based Analysis and Retrieval of Explanations), an interactive system that treats global explanations not as static documents but as a queryable database.

System Architecture

The system follows a parse-validate-execute pipeline:

1. Input: Users pose natural language questions regarding aggregate model behavior (e.g., "What percentage of bedroom images contain both bed and wall?").
2. Semantic Parsing (LLM): A fine-tuned Large Language Model (LLM) acts as a semantic parser. It translates the user's natural language query into a structured SQL query.
   • The LLM selects from a predefined taxonomy of 24 query templates covering tasks like frequency counts, boolean combinations, co-occurrence, and counterfactual analysis.
   • It extracts parameters (class names, object names, thresholds) to instantiate the template.
3. Execution: The generated SQL is validated for safety and correctness, then executed against a relational database containing the global explanation data.
   • The database is constructed by aggregating Local Minimal Sufficient Explanations (MSXs) into a global Disjunctive Normal Form (DNF) structure.
   • While the experiments use explanations generated by a specific method (Vasu et al. [20]), the interface is designed to work with any concept-based local explanation method that produces symbolic forms.
4. Output: The system returns:
   • A natural language summary augmented with relevant statistics.
   • Supporting local explanations (evidence images).
   • Visualizations aligned with the query intent (e.g., segmentation maps highlighting relevant objects).

Training Strategy

To ensure the LLM learns the structure of explanation querying rather than memorizing specific dataset entities, the authors employ a specialized fine-tuning approach:

• Synthetic Data: 50,000 training pairs are generated by sampling random combinations of objects and scene classes from the ADE20K vocabulary and applying natural language variations (synonyms, phrasing templates).
• SQL-Fence Loss Masking: During fine-tuning (using LoRA on Gemma 2 and Qwen 2.5 models), the training loss is masked to focus exclusively on tokens between SQL_START and SQL_END markers. This forces the model to learn the relational algebra of the query language, enabling zero-shot transfer to new datasets with different vocabularies.

3. Key Contributions

The paper outlines three primary contributions:

1. Natural Language Interface for Global Explanations: The introduction of GLARE, which transforms global explanations into queryable databases, allowing users to interrogate model behavior via natural language.
2. SQL-Based Intermediate Representation: The definition of a structured SQL schema for aggregating, filtering, and contrasting local explanations, bridging the gap between natural language intent and formal logical aggregation.
3. Robust Query Interpretation via Synthetic Fine-Tuning: Demonstration that synthetic-data fine-tuning with SQL-fence loss masking yields robust query interpretation. This approach enables zero-shot cross-dataset transfer (e.g., from ADE20K to Pascal VOC) without retraining, as the model learns generalized reasoning patterns rather than dataset-specific entity mappings.

4. Experimental Results

The system was evaluated on global explanations derived from the ADE20K scene parsing dataset, with a cross-dataset evaluation on Pascal VOC.

In-Distribution Accuracy

• Performance: Fine-tuned models (specifically Gemma 2 9B and 27B) achieved >95% result-match accuracy on held-out queries over the training dataset.
• Structural Integrity: These models achieved near-perfect scores on fence detection, SQL parsing, and execution rates (100% in many cases).
• Baseline Comparison: Base (untrained) models achieved near-zero accuracy, confirming that the performance is a product of the specific fine-tuning pipeline rather than pre-existing SQL knowledge. A non-learning Regex baseline achieved 74.4% accuracy, significantly lower than the fine-tuned models.

Robustness to Perturbations

The system demonstrated strong resilience to linguistic noise:

• Spelling Errors: The fine-tuned model maintained a 79.6% match rate compared to the Regex baseline's 48.7%, showing a +31 percentage point advantage.
• Synonym Substitution: The model exhibited 100% robustness to synonyms, outperforming the Regex baseline by 17 percentage points.
• Consistency: Semantically equivalent paraphrases yielded identical SQL results 94.1% of the time.
• Sanity: All outputs satisfied domain constraints (e.g., no negative counts or percentages outside [0, 100]).

Generalization and Transfer

• Out-of-Distribution (OOD) Phrasing: The model handled novel phrasing of trained templates well (45.1% exact match), though it struggled with complex double negations or highly informal queries.
• Novel SQL Constructs: The model showed compositional generalization, assembling familiar SQL fragments into new structures (e.g., chained_filter at 100% match) but failed on entirely unseen syntax (e.g., window_function at 0% match).
• Cross-Dataset Transfer: A model trained exclusively on ADE20K metadata achieved ~90% accuracy when tested on a Pascal VOC database with a completely disjoint object vocabulary. This confirms the model learned the relational structure of the schema rather than specific entity names.

Model Scale Analysis

• Capacity Threshold: A minimum model capacity is required. Qwen 2.5 0.5B failed to learn the task even after fine-tuning, while 7B+ models performed robustly.
• Saturation: Within the Gemma 2 family, performance saturated across 2B, 9B, and 27B parameters for in-distribution tasks, suggesting the synthetic pipeline is effective even for smaller, sufficiently large models.

5. Significance and Claims

The paper claims that GLARE offers a practical and reliable path toward human-centered XAI. By shifting from passive inspection of static artifacts to active, query-driven dialogue, the system makes global explanations accessible to non-experts.

The authors emphasize that their approach:

• Grounds Flexibility in Formal Correctness: By using LLMs as logic-constrained parsers mapping to verifiable SQL templates, the system avoids the hallucinations common in direct LLM-based explanation generation.
• Enables Zero-Shot Transfer: The "SQL-fence" training strategy allows the system to generalize to new domains and vocabularies without retraining, provided a new entity name list is available.
• Reduces Cognitive Load: It allows users to ask targeted questions without needing to understand the underlying complexity of thousands of logical rules.

The paper concludes that while the current template taxonomy is extensible but not exhaustive, the demonstrated ability to transfer across datasets and handle linguistic variations suggests that LLM-mediated querying is a viable solution for scaling global explanation analysis.