Workman et al. (scBench team)

SpatialBench: Can Agents Analyze Real-World Spatial Biology Data?

146 verifiable problems across five spatial technologies and seven task categories, from the scBench team. Each gives a data snapshot just before an analysis step plus a deterministic grader for recovering the key biological result. Accuracy is low (20–38% across families) and very sensitive to harness design — squarely in the spatial-omics wheelhouse. Seven frontier models evaluated; Claude Opus 4.5 leads at 38.4%.

spatial-biologyspatial-transcriptomicstabularimagecode
Status
cataloged
Saturation
unsaturated
Runnable
no
Cost to run
high

Editor's note

146 verifiable problems across five spatial technologies and seven task categories, from the scBench team. Each gives a data snapshot just before an analysis step plus a deterministic grader for recovering the key biological result. Accuracy is low (20–38% across families) and very sensitive to harness design — squarely in the spatial-omics wheelhouse. Seven frontier models evaluated; Claude Opus 4.5 leads at 38.4%.

Results

ModelScoreVariantProvenance
Claude Opus 4.538.4% accuracydefaultpaper
GPT-5.234.0% accuracydefaultpaper
Claude Sonnet 4.528.3% accuracydefaultpaper
GPT-5.127.4% accuracydefaultpaper
Grok-4.124.7% accuracydefaultpaper
Grok-422.8% accuracydefaultpaper
Gemini 2.5 Pro20.1% accuracydefaultpaper

Sources

  • paper SpatialBench (arXiv:2512.21907)