SpaceDG: Benchmarking Spatial Intelligence under Visual Degradation

Researchers introduced SpaceDG, a degradation-aware dataset and benchmark for spatial reasoning with about 1 million QA pairs and 10,000+ VQA examples across nine visual degradation types.

Testing 25 multimodal models showed large drops in spatial intelligence when inputs were blurred, darkened, compressed, or affected by adverse weather.

Fine-tuning on SpaceDG substantially improved degraded-image performance and even surpassed human results in degraded settings.

The gains did not come at the cost of clean-image accuracy, highlighting a practical path to more robust multimodal AI.