GridVQA-X: Diagnostic Framework for Evaluating Cross-Modal Explainers

GridVQA-X is a diagnostic framework for evaluating the faithfulness of post-hoc cross-modal explainers. It features S × S visual grids populated by geometric objects paired with questions, using a closed-world synthesis logic with mathematically guaranteed unique ground-truth explanations. The dataset was created by Aikyam-Lab and was last updated on Hugging Face in June 2026.

Use Cases

• Benchmarking the faithfulness of visual question answering (VQA) explainers based on the controlled synthetic environment.
• Isolating genuine cross-modal spatial reasoning from shallow shortcuts based on the closed-world synthesis logic.
• Developing new evaluation metrics for post-hoc explainability based on the mathematically guaranteed unique ground-truth explanations.

Strengths

• Provides a controlled testbed with mathematically guaranteed unique ground-truth explanations, enabling objective evaluation.
• Designed as the first diagnostic framework specifically for post-hoc cross-modal explainers, filling a methodological gap.

Limitations

• Description metadata is limited; actual data quality, scale, and specific file formats require manual inspection after download.
• Row count, column details, and sample data are unknown, which may limit suitability assessment for specific tasks.

Provenance

Source

Aikyam-Lab via Hugging Face

Collection Method

Synthesized using a closed-world logic with geometric objects on visual grids.

Freshness

Last updated 2026-06-24 15:41:51; freshness should be verified.