OmniSpatial: A Visual-Spatial Reasoning Benchmark for Vision Language Models

OmniSpatial is a benchmark dataset for evaluating spatial reasoning in vision-language models, as presented in an ICLR 2026 paper. The data is structured in a JSON schema with components like 'id' for question identification. The dataset was created by author 'qizekun' and last updated on January 27, -2026.

Use Cases

• Benchmarking spatial reasoning performance of vision-language models based on the structured task schema.
• Training models for visual question answering tasks that involve spatial relationships.
• Analyzing model weaknesses in multimodal spatial understanding based on the benchmark's question-answer pairs.

Strengths

• Designed for a specific research purpose, presented at a major conference (ICLR 2026).
• Includes a documented JSON task schema for structured evaluation.

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.
• Row count, file formats, and sample data are unknown, which may limit suitability assessment.

Provenance

Source

huggingface

Collection Method

Created for research presented at ICLR 2026.

Time Range

null

Freshness

Last updated 2026-01-27 09:09:53; freshness should be verified.

Geography

null