SpatialBlock: 15K synthetic puzzles train VLMs on 3D reasoning

A Qwen3-VL-4B model fine-tuned on 15,000 synthetic block-stacking puzzles scored 95.7% on its in-domain benchmark and 51.3% on the out-of-domain MindCube test, according to a new arXiv paper from Soohyun Ryu, Sohee Kim, and Eunho Yang.

The dataset, SpatialBlock-15k, covers "3D-to-2D projection, viewpoint transformation, and structural combination," per the abstract, and every example is procedurally generated, sidestepping the dense human-labelled geometry that most spatial-reasoning datasets rely on. The authors describe their bet as "learning foundational spatial skills through structured block-manipulation tasks," an approach the abstract says is "inspired by human cognitive development."

On out-of-domain benchmarks, the 7B variant reached 50.9% on SPBench, 29.7% on MMSI-Bench, and 55.5% on MMMU. Human evaluators average 95% accuracy on the in-domain SB-Bench tasks, the paper reports. The authors argue that LVLMs trained on this data "significantly outperform baselines and generalize to real-world spatial tasks, despite the dataset's synthetic and compact nature."

The abstract publishes no head-to-head against closed-source frontier models, and the load-bearing claim is how far a compact block-stacking curriculum actually transfers to messy 3D scenes rather than benchmarks tilted toward geometric primitives. Code and data, the authors say, are on the repository linked from the paper.