VQA-G Annotator: A Self-Improving Agentic Framework for High-Fidelity Grounded Dataset Synthesis

As artificial intelligence systems increasingly rely on multi-source perception and cross-modal learning for human-like visual understanding, their outputs should be both semantically valid and spatially grounded. However, large-scale vision-language benchmarks with precise grounding remain costly to construct, while existing synthetic pipelines often suffer from distribution drift, error propagation, and weak spatial verification. We introduce VQA-G Annotator, a self-improving agentic framework that combines distribution-aware planning, multi-agent orchestration, and spatial reasoning verification in a closed-loop generation, evaluation, and refinement process. Experiments across four benchmark sources show that VQA-G Annotator achieves an average VQAScore of 0.88 for semantic alignment and an Acc@0.5 of 0.57 for spatial grounding, comparing favorably with strong automatic baselines. Downstream and human evaluations further support the utility of the synthesized data 18 for scalable and trustworthy grounded visual understanding.