BioVQA: A Benchmark for Biological Question Answering across Taxonomic, Fine-grained and Spatial Reasoning
IEEE/CVF Winter Conference on Applications of Computer Vision (WACV 2027), Datasets and Benchmarks track
Can a VLM name the genus, read the marked wing, count the specimens and tell two zebras apart? BioVQA tests each ability on its own, across 1.87M questions.
Vision-language models are now asked to read biological images, from field-guide style identification to measuring specimens in a collection. A single accuracy number hides which of these abilities a model has. A model can name a species from text priors without looking, or describe a bird well and still confuse its left and right wing.
BioVQA splits the problem into four components with their own data and metrics, and places every model between an image-blind floor, which answers without the image, and a specialist ceiling. The gap to the floor shows how much a model actually uses the image.
Four components
Hierarchical taxonomic reasoning
Questions from kingdom down to species on TreeOfLife-200M images across 10 organism groups. Answers are checked against GBIF synonyms, and BioCLIP-2 serves as the specialist ceiling.
Fine-grained attribute grounding
824K questions on CUB-200-2011 about the color, pattern and location of marked body parts, scored with the CAGS metric and checked by three human annotators. Laterality checks catch left-right confusions.
Quantitative spatial reasoning
Counting, viewpoint, relative and absolute distance, and OCR on specimen trays, where answers can be verified exactly.
Individual re-identification
Same-or-different questions about individual animals for 5 species under 3 conditions, scored with signal detection theory (d′ and criterion c), so a model cannot win by always answering yes.
Evaluation
| Scale | 1.87M questions over 216K images |
|---|---|
| Models | 18 open-weight VLMs from 3.75B to 38.4B parameters, including Qwen3-VL, InternVL3.5, Gemma 3 and 4, LLaVA-NeXT, Llama 3.2 Vision, Pixtral, Molmo2 and Phi-4-multimodal |
| Decoding | Greedy decoding with short, scorable answers |
| Reference points | Image-blind floor for every task; specialist models as the ceiling |
| Infrastructure | SLURM array pipelines on A100 and H100 clusters at the Ohio Supercomputer Center |
Why it matters
Biologists will only trust a VLM for collection work if they know which kind of question it can answer. BioVQA gives a per-ability profile instead of a single leaderboard position, so a lab can choose a model for counting specimens without assuming it can also identify them.
The benchmark and evaluation code are on GitHub.
Related work
Identifying and Resolving Pitfalls of Knowledge-Based VQA Benchmarks: Auditing, Repairing, and Augmenting
European Conference on Computer Vision (ECCV 2026)Fine-Grained Beetle Taxonomy with Vision Models: A Benchmark on Long-Tailed and Domain-Adaptive Classification
ICCV 2025 Workshop on Computer Vision for Ecology (CV4E)Mind the (Data) Gap: Evaluating Vision Systems in Small Data Applications
NeurIPS 2025 Workshop on Imageomics
Cite
@inproceedings{rayeed2027biovqa,
title = {BioVQA: A Benchmark for Biological Question Answering across Taxonomic, Fine-grained and Spatial Reasoning},
author = {Rayeed, S M and others},
booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
year = {2027}
}