SPaRTA: Spatial Reasoning in Text-to-SQL for Omics Analysis

SPaRTA-Bench shows open-weight LLMs achieve 10–40% zero-shot execution accuracy for spatial Text-to-SQL over omics tissue sections, rising to 63–77% after fine-tuning.

• Why it matters: Spatial-omics queries are hard for non-programmers, and existing systems lack validated benchmarks that expose where models fail when translating natural language into spatial SQL.
• What they did: The team built SPaRTA to convert free-text into executable spatial SQL, released SPaRTA-Bench with 7,500 validated question–SQL pairs across breast cancer, pancreatic cancer, and melanoma, and cataloged 15 wrong-answer failure modes.
• The result: Reasoning plus execution feedback reaches up to 61% and fine-tuning improves to 63–77%, with at most a 6-point drop on an unseen tissue; feedback mainly fixes run-time failures, not wrong answers.

AI summary of the abstract · may contain errors · not medical advice · read the paper.