Exploring Spatial Intelligence from a Generative Perspective
Summary
Researchers introduce GSI-Bench, the first benchmark to quantify generative spatial intelligence in multimodal models by evaluating 3D spatial constraint compliance during image generation. Fine-tuning on their synthetic dataset boosts both spatial editing fidelity and downstream spatial understanding, showing generative training can strengthen spatial reasoning.
View Cached Full Text
Cached at: 04/23/26, 07:47 AM
Paper page - Exploring Spatial Intelligence from a Generative Perspective
Source: https://huggingface.co/papers/2604.20570 Authors:
,
,
,
,
,
,
,
,
,
Abstract
Generative spatial intelligence benchmark evaluates and enhances 3D spatial constraint manipulation in image generation through real-world and synthetic datasets.
Spatial intelligence is essential formultimodal large language models, yet current benchmarks largely assess it only from an understanding perspective. We ask whether modern generative or unified multimodal models also possessgenerative spatial intelligence(GSI), the ability to respect and manipulate3D spatial constraintsduringimage generation, and whether such capability can be measured or improved. We introduceGSI-Bench, the first benchmark designed to quantify GSI throughspatially grounded image editing. It consists of two complementary components:GSI-Real, a high-quality real-world dataset built via a3D-prior-guided generationand filtering pipeline, andGSI-Syn, a large-scale synthetic benchmark with controllable spatial operations and fully automated labeling. Together with a unified evaluation protocol,GSI-Benchenables scalable,model-agnostic assessmentofspatial complianceandediting fidelity. Experiments show thatfine-tuningunified multimodal models onGSI-Synyields substantial gains on both synthetic and real tasks and, strikingly, also improves downstream spatial understanding. This provides the first clear evidence that generative training can tangibly strengthenspatial reasoning, establishing a new pathway for advancing spatial intelligence in multimodal models.
View arXiv pageView PDFProject pageGitHub5Add to collection
Get this paper in your agent:
hf papers read 2604\.20570
Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash
Models citing this paper0
No model linking this paper
Cite arxiv.org/abs/2604.20570 in a model README.md to link it from this page.
Datasets citing this paper0
No dataset linking this paper
Cite arxiv.org/abs/2604.20570 in a dataset README.md to link it from this page.
Spaces citing this paper0
No Space linking this paper
Cite arxiv.org/abs/2604.20570 in a Space README.md to link it from this page.
Collections including this paper0
No Collection including this paper
Add this paper to acollectionto link it from this page.
Similar Articles
VGenST-Bench: A Benchmark for Spatio-Temporal Reasoning via Active Video Synthesis
VGenST-Bench is a benchmark that uses generative models to actively synthesize controlled spatio-temporal reasoning scenarios, with a multi-agent pipeline and human quality control, to evaluate multimodal large language models.
ESI-Bench: Towards Embodied Spatial Intelligence that Closes the Perception-Action Loop
Introduces ESI-BENCH, a comprehensive benchmark for embodied spatial intelligence built on OmniGibson, covering 10 task categories and 29 subcategories. Experiments show active exploration substantially outperforms passive approaches, with failures mainly due to action blindness rather than perception, revealing a metacognitive gap in models compared to humans.
SpatialWorld: Benchmarking Interactive Spatial Reasoning of Multimodal Agents in Real-World Tasks
SpatialWorld is a unified benchmark for evaluating interactive spatial reasoning in multimodal agents across diverse real-world tasks, revealing that even the strongest models achieve low task success rates.
SpatialBench: Is Your Spatial Foundation Model an All-Round Player?
SpatialBench is a comprehensive benchmark for evaluating spatial foundation models across diverse domains and tasks, revealing limitations in current models and introducing DA-Next-5M and DA-Next to advance spatial representation learning.
OVO-S-Bench: A Hierarchical Benchmark for Streaming Spatial Intelligence in Multimodal LLMs
OVO-S-Bench introduces a comprehensive human-annotated benchmark of 1,680 questions across 348 videos to evaluate streaming spatial intelligence in multimodal LLMs, revealing that even the best model (Gemini-3.1-Pro) trails human experts by 27 points. The benchmark exposes key limitations including allocentric mapping as a major bottleneck and chain-of-thought reasoning amplifying spatial errors.