SpatialBlock: Enhancing Spatial Intelligence in LVLMs via Synthetic Block-Stacking Problem
Soohyun Ryu, Sohee Kim, Eunho Yang
KAIST · AITRICS
Published September 7, 2026 · on the feed September 11, 2026 · ▲ 143 · arXiv 2609.07064
The abstract
Large Vision-Language Models (LVLMs) have achieved strong performance on diverse visual tasks, yet their ability to reconstruct and reason about the 3D structure of the scene depicted in 2D images -- referred to as spatial intelligence -- remains limited. Existing approaches attempt to address this gap by using real-scene spatial question answering datasets that require dense geometric annotations. However, constructing such labels is costly, time-consuming, and often noisy due to reliance on external perception modules. In this work, we propose a novel paradigm inspired by human cognitive development: learning foundational spatial skills through structured block-manipulation tasks. We introduce SpatialBlock-15k, a synthetic dataset of 15,000 block-stacking problems covering 3D-to-2D projection, viewpoint transformation, and structural combination. The dataset further incorporates controlled color modulation as visual cues to encourage anchor-based reasoning in visually complex conditions. Experiments demonstrate that LVLMs trained on our dataset through either direct answering or reasoning-based prediction significantly outperform baselines and generalize to real-world spatial tasks, despite the dataset's synthetic and compact nature. Code and data are available at https://github.com/rsoohyun/SpatialBlock.
The Brief
2:14A short podcast summary of the paper. Written and voiced by AI.
Transcript
SpatialBlock. Enhancing Spatial Intelligence in LVLMs via Synthetic Block-Stacking Problem. By Soohyun Ryu and two others, from KAIST and AITRICS. Published on the seventh of September. Its code repository is worked out: about seventy-five source files, with a readme and a licence.
Large vision-language models, LVLMs, do well on a wide range of visual tasks. What remains limited is what the authors call spatial intelligence: reconstructing, and reasoning about, the 3D structure of a scene from a 2D image of it. SpatialBlock-15k is a dataset built to teach that skill by a different route. Existing approaches rely on real-scene spatial question answering. This one borrows from human cognitive development instead, and trains foundational spatial skills through structured tasks of manipulating blocks. The claim is that a model can learn the basics of seeing in three dimensions from a toy world, and then carry them into the real one.
It holds fifteen thousand synthetic block-stacking problems. They cover three kinds of skill. The first is projecting from 3D to 2D. The second is transforming between viewpoints. The third is structural combination, putting pieces together into a whole. The dataset also includes controlled colour modulation, used as a visual cue to encourage the model to reason from anchors when the visual conditions are complex. Each problem is a small constructed scene rather than a photographed one.
The problems are synthetic, and that is deliberate. Real-scene spatial datasets need dense geometric annotations, and those labels are costly, time-consuming to produce, and often noisy, because they depend on external perception modules. Generating block problems sidesteps that labelling pipeline entirely, and keeps the dataset compact.
The dataset is for anyone training LVLMs toward spatial reasoning. In the experiments, models trained on it outperform baselines significantly, and this holds for either training route: direct answering, or prediction that goes through reasoning. They also generalize to real-world spatial tasks, despite the data being synthetic and small. Code and data are publicly available.
The bet resembles how a child learns: stack blocks before reasoning about rooms. A small, cleanly labelled world of blocks carries skills that transfer to real scenes, without anyone having to annotate the geometry of a real one.
Paper Quest · 7 chapters · about 41 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
Models that ace 2D images still cannot picture the room behind them. That is the wall.
You want a vision-language model that can look at a 2D photo and work out the 3D layout: what is behind what, and how the scene looks from the other side.
The road ahead
- 1 · The problem ~7 min Vision-language models read flat photos well but struggle to rebuild the 3D scene behind them, and the usual training data for that is expensive and noisy.
- 2 · What came before ~3 min
- 3 · The key idea ~9 min Teach spatial basics with generated block-stacking puzzles, whose labels are clean and free, and use colours as anchors for reasoning.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Generate three kinds of block puzzles, add colour cues, train models to answer directly or by reasoning, then test on real-world spatial benchmarks.
- 6 · The experiment ~6 min You’ll guess: Trained only on compact synthetic block puzzles, did the LVLMs carry their gains over to real-world spatial tasks?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.