| --- |
| language: |
| - en |
| license: mit |
| task_categories: |
| - image-text-to-text |
| - visual-question-answering |
| tags: |
| - vision-language |
| - vqa |
| - diagnostic-benchmark |
| - xai |
| --- |
| |
| # GridVQA-X Datasets |
|
|
| GridVQA-X is the first diagnostic framework designed to objectively evaluate the faithfulness of post-hoc cross-modal explainers. By utilizing a closed-world synthesis logic with mathematically guaranteed unique ground-truth explanations, it provides a controlled testbed to isolate genuine cross-modal spatial reasoning from shallow shortcuts. |
|
|
| - **Paper:** [GridVQA-X: A Framework for Evaluating Multimodal Explainability Methods](https://huggingface.co/papers/2606.14740) |
| - **Repository:** [GitHub - AikyamLab/grid-vqax](https://github.com/AikyamLab/grid-vqax) |
|
|
| ## Dataset Summary |
| The dataset features **S × S** visual grids populated by geometric objects, paired with multi-hop spatial reasoning queries. Each query defines a **Target** object to find/count and one or more **Anchor** reference objects connected by directional tokens. |
|
|
| The dataset contains two distinct, parallel splits: |
| * **D_pure (The Causal Split):** Systematically removes all statistical heuristics. It algorithmically populates invalid spatial regions with adversarial distractors matching target attributes, forcing models to utilize complete multi-hop spatial geometry. |
| * **D_spur (The Shortcut Split):** Intentionally preserves the pervasive "Bag-of-Words Alignment" cross-modal shortcut. No target-attribute objects are placed outside the valid spatial region, creating a behavioral trap where the probability of the answer given target attributes is 1.0. |
|
|
| ## Dataset Taxonomy and Configurations |
| Every sample is parameterized by a strict 4-tuple configuration: **(Depth, QType, Form, Density)** |
|
|
| * **Spatial Depth (1, 2, or 3):** Controls the number of anchor objects and relational multi-hop complexity. |
| * **Question Type (QType):** |
| * `Attribute-Only (A)`: Non-spatial baseline queries. |
| * `Shape-Only (SO)` / `Color-Only (CO)`: Isolates directional reasoning to a single feature type. |
| * `Mixed (M)`: Requires multimodal binding of shape and color for both target and anchor. |
| * `Comparison (CMP)`: Multi-step logical operations comparing object counts. |
| * **Form (0 or 1):** Alters the target task header between counting (`0`) and existence (`1`) to prevent statistical format overfitting. |
| * **Density (d_0.3 or d_0.7):** Adjusts grid object count to test spatial localization fidelity against background noise. |
|
|
| ## Ground-Truth Explanations |
| The dataset provides a mathematically proven, unique ground-truth causal explanation. The ground-truth visual masks are strictly bounded to the target and anchor items (**A ∪ T**), leaving all other distractor objects with a causal effect of exactly zero. |
|
|
| ## Citation |
|
|
| If you use GridVQA-X in your research, please cite: |
|
|
| ```bibtex |
| @misc{belsare2026gridvqaxframeworkevaluatingmultimodal, |
| title={GridVQA-X: A Framework for Evaluating Multimodal Explainability Methods}, |
| author={Sujay Belsare and Sudarshan Nikhil and Sushant Kumar and Ponnurangam Kumaraguru and Chirag Agarwal}, |
| year={2026}, |
| eprint={2606.14740}, |
| archivePrefix={arXiv}, |
| primaryClass={cs.CV}, |
| url={https://arxiv.org/abs/2606.14740}, |
| } |
| ``` |
