PlantMarkerBench: A Multi-Species Benchmark for Evidence-Based Gene Marker Reasoning

This dataset is part of an anonymous submission to NeurIPS 2026.

PlantMarkerBench is a multi-species benchmark for evaluating evidence-based gene marker reasoning from plant biology literature. The dataset is constructed using an automated pipeline that integrates large-scale literature retrieval, hybrid search, and structured evidence extraction, followed by human verification.

The dataset spans four plant species: Arabidopsis, maize, rice, and tomato. Each instance consists of a gene, cell type, and candidate evidence sentence, annotated with:- validity of marker evidence- evidence type (expression, function, localization, indirect, noise)- support strength

We define a reasoning task where models must determine whether a given sentence supports a gene as a valid cell-type marker and classify the type of evidence.

For review purposes, we provide a representative sample (n=600) with balanced valid/invalid examples and diverse evidence types. The full dataset is also included.

This dataset is intended for benchmarking scientific reasoning in language models and not for direct biological decision-making without expert validation.