GroundVQA

GroundVQA is an adversarial visual question answering (VQA) benchmark for evaluating visual grounding, hallucination resistance, uncertainty calibration, OCR robustness, and physically grounded reasoning in frontier multimodal large language models (MLLMs).

The current release contains approximately 4,000 manually reviewed adversarial VQA examples constructed from images of authors and multiple public image sources of Visual Genome, WearVQA, TextCaps, DocVQA, ChartQA, and GQA (QA re-curated).

Dataset Description

GroundVQA focuses on evaluation rather than training. Compared with conventional VQA benchmarks, many examples require models to:

• reject unsupported premises
• abstain when visual evidence is insufficient
• distinguish visible observations from inferred content
• avoid hallucinating objects, attributes, relationships, or OCR text

The benchmark covers multiple reasoning categories including:

• OCR grounding
• Spatial reasoning
• Physical reasoning
• Adversarial hallucination
• Document understanding
• Chart reasoning
• Ambiguous visual evidence

Intended Uses

GroundVQA is intended for:

• Evaluation of frontier MLLMs
• Hallucination analysis
• Visual grounding research
• Benchmarking uncertainty-aware multimodal reasoning

It is not intended as a supervised training dataset.

Citation

(The citation will be updated after publication.)