MMSI-Bench: A Benchmark for Multi-Image Spatial Intelligence

MMSI-Bench is an evaluation dataset for multi-image spatial intelligence tasks, created by RunsenXu. The dataset was last updated on October 23, 2025, and includes normalized human response times and difficulty levels for its samples. It is referenced in a research paper and supported by the LMMs-Eval repository.

Use Cases

• Benchmarking model performance on multi-image spatial reasoning tasks based on the dataset's evaluation framework.
• Analyzing the correlation between human response time and task difficulty for spatial intelligence problems.
• Training or fine-tuning vision-language models on tasks requiring understanding of spatial relationships across images.

Strengths

• Includes normalized human response time data for each sample, providing a human performance baseline.
• Each sample has an associated difficulty level, allowing for stratified analysis of model performance.
• Dataset was updated on 2025-10-23, indicating recent maintenance.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• The description metadata is limited; actual data quality and content require manual inspection after download.

Provenance

Source

RunsenXu

Collection Method

Created as an evaluation benchmark for a research paper on multi-image spatial intelligence.

Freshness

Last updated 2025-10-23 13:04:41.