SpaceDG: Benchmarking Spatial Intelligence under Visual Degradation

A new dataset called SpaceDG reveals that visual degradations such as motion blur and low light consistently impair the spatial reasoning of multimodal large language models, exposing a critical gap between pristine benchmarks and real-world performance, according to a study published on arXiv [1][2]. The study introduces SpaceDG, described as the first large-scale dataset for degradation-aware spatial understanding [1][2]. It contains approximately 1M QA pairs drawn from nearly 1,000 indoor scenes [1][2]. The dataset was built using a physically grounded degradation synthesis engine that embeds degradation formation into 3D Gaussian Splatting rendering, enabling realistic simulation of nine degradation types including motion blur, low light, adverse weather, lens distortion, and compression artifacts [2]. Compression artifacts are a common byproduct of lossy compression, a process that reduces file size by removing information deemed less important [3]. Alongside the dataset, the researchers constructed SpaceDG-Bench, a human-verified benchmark comprising 1,102 questions across 11 reasoning categories and the same nine degradation types, yielding over 10,000 visual question-answering instances [1][2]. The team evaluated 25 open- and closed-source MLLMs on the benchmark [1][2]. The results showed that visual degradations consistently and substantially impair spatial reasoning [1][2]. Large language models are machine learning models with many parameters, trained with self-supervised learning on vast amounts of text [7]. Multimodal variants extend this capability to visual inputs. The SpaceDG study found that finetuning models on the new dataset markedly improved their robustness to degradation [1][2]. In some cases, the finetuned models surpassed human performance under degraded conditions, without any drop in accuracy on clean images [1][2]. The authors argue this highlights the promise of degradation-aware training for robust spatial intelligence [1][2]. The work was submitted to arXiv on May 21, 2026, with a revised version posted on June 26, 2026, by Xiaolong Zhou and collaborators [1].