| --- |
| language: |
| - en |
| license: mit |
| pipeline_tag: image-text-to-text |
| tags: |
| - vision-language |
| - mdetr |
| - xai |
| model_index: |
| - name: mdetr-gridvqa-pure |
| task: visual-question-answering |
| - name: mdetr-gridvqa-spurious |
| task: visual-question-answering |
| --- |
| |
| # GridVQA-X Models |
|
|
| This repository contains two paired reference models, **M_pure** and **M_spur**, built on identical transformer architectures (**MDETR**). These models, coupled with their corresponding datasets, together form a diagnostic framework to evaluate if Multimodal Explainable AI (MxAI) methods genuinely capture cross-modal synergy or simply report shallow feature correlations. |
|
|
| This model is presented in the paper [GridVQA-X: A Framework for Evaluating Multimodal Explainability Methods](https://huggingface.co/papers/2606.14740). |
| The official training and evaluation code can be found in the [GitHub Repository](https://github.com/AikyamLab/grid-vqax). |
|
|
| ## Model Descriptions |
|
|
| ### 1. M_pure (The Faithful Spatial Reasoner) |
| * **Training Distribution:** Trained exclusively on the D_pure dataset. |
| * **Behavioral Dynamics:** Trained via explanation-guided dynamics using a two-phase optimization. Phase 1 forces explicit visual-textual token alignment via L1 and generalized IoU losses. Phase 2 handles Question Answering using class-frequency weighted cross-entropy to completely eliminate answer prior biases. |
| * **Capabilities:** Successfully internalizes true causal spatial-relational synergy, achieving robust accuracy across both clean and heavily distractor-crowded grids. |
|
|
| ### 2. M_spur (The Shortcut / Bag-of-Words Model) |
| * **Training Distribution:** Trained exclusively on the D_spur dataset. |
| * **Behavioral Dynamics:** Structurally forced to rely on cross-modal shortcuts during training. It skips relational spatial geometry entirely and maps keywords directly to target visual volume. |
| * **Capabilities:** Achieves perfect accuracy (1.000) on its native spurious distribution, but fails catastrophically when evaluated on D_pure multi-hop queries. |
| |
| ## Intended Diagnostic Use |
| These models are released explicitly to stress-test vision-language explainability algorithms (e.g., DIME, MultiSHAP, MultiViz, EMAP, InterSHAP): |
| * **The Litmus Test:** A faithful explainer *must* output completely different attribution heatmaps or synergy scalars for M_pure and M_spur on the same input question. |
| * **The Reality Check:** If your explainer highlights identical spatial regions for both models, it suffers from "model blindness" or is simply behaving as a superficial object detector. |
| |
| ## Performance Benchmark Metrics |
| |
| | Evaluation Metric | M_pure on D_pure | M_spur on D_spur | M_spur on D_pure | |
| | :--- | :---: | :---: | :---: | |
| | **Global Accuracy** | >99% | 100% | **Catastrophic Failure** (8%-14% on multi-hop) | |
| | **Causal Pathway** | True Spatial Relations | Bag-of-Words Shortcut | Unimodal Feature Collapse | |
| |
| ## Citation |
| |
| ```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}, |
| } |
| ``` |
