How Finely Can Vision–Language Models See Biology?
Bio-VQA: A Benchmark for Taxonomic, Morphological, Spatial and Individual Reasoning

S M Rayeed · Qian Ma · Charles V. Stewart · Yao Ma

🌐 Project page  |  💻 Github Repo  |  📄 Paper  |  ⚡ Bio-VQA-Lite

---

Dataset summary

Bio-VQA is a biological visual question answering benchmark: 1,861,608 questions over 216,715 images drawn from 11 public datasets, organized into four components and 24 tasks. Together they cover four aspects of biological visual understanding that prior benchmarks treat separately, or not at all:

	Component	What it asks	Tasks	Images	Questions	Source datasets
C1	Hierarchical taxonomic reasoning	Which taxon, at eight ranks?	9	100,000	890,000	TreeOfLife-200M
C2	Fine-grained attribute grounding	Which trait, on which part, on which side?	4	23,576	824,104	CUB-200-2011
C3	Quantitative spatial reasoning	How many, how far, where, facing where?	8	78,139	132,504	WAID, Boxes on Bees, 2018-NEON-Beetles, DePondFi, GZCD, WhaleSharkID, HyenaID, LeopardID
C4	Individual re-identification	Is this the same animal?	3	15,000	15,000	WhaleSharkID, BelugaID, HyenaID, LeopardID, GZCD
	Total		24	216,715	1,861,608	11 (four serve both C3 and C4)

Three design rules run through the whole benchmark:

1. Ground truth comes only from existing annotations. A question is generated only where a source annotation (lineage, attribute, box, count, measurement, viewpoint, individual ID) supplies the answer. No answer key depends on our own labeling, so every answer can be checked against its source.
2. Every task has a reference point. Each task carries an image-blind floor (the best a rule that never looks at the image can score), and a domain specialist ceiling where one exists (BioCLIP-2 for C1, MegaDescriptor for C4).
3. Graded scoring. A near miss scores higher than a distant miss: Wu–Palmer and Lin similarity for C1, CAGS for C2, CAS and lift over the floor for C3, d′ and criterion c for C4. Each graded score is reported next to the exact match it modifies.

This card describes the data. The question generators, the 18 model runners and the scorers are in the GitHub repository. Figures and examples are on the project page.

---

Contents

• Quick start
• Repository layout
• Configurations and splits
• Components in detail
• Bio-VQA-Lite
• Data fields
• Images
• Model predictions
• How the data were built
• Scoring
• Results summary
• Intended use and limitations
• Licenses
• Citation

---

Quick start

from datasets import load_dataset

lite = load_dataset("imageomics/Bio-VQA", "lite")                  # splits: C1, C2, C3, C4 (12,000 questions)
c3   = load_dataset("imageomics/Bio-VQA", "C3_spatial", split="counting")
c4   = load_dataset("imageomics/Bio-VQA", "C4_reidentification", split="hard")

Images for C2 marker localization, C3 and C4 are stored in zip shards next to the question files. Each row names its shard (image_zip) and the file inside it (image_in_zip):

import io, zipfile
from huggingface_hub import hf_hub_download
from PIL import Image

def load_image(component, image_zip, image_in_zip):
    """component: 'C2', 'C3' or 'C4'"""
    path = hf_hub_download("imageomics/Bio-VQA", f"{component}/{image_zip}", repo_type="dataset")
    with zipfile.ZipFile(path) as z:
        return Image.open(io.BytesIO(z.read(image_in_zip))).convert("RGB")

row = c4[0]
img = load_image("C4", row["image_zip"], row["image_in_zip"])

To download a whole component at once (for example C3):

from huggingface_hub import snapshot_download
snapshot_download("imageomics/Bio-VQA", repo_type="dataset", allow_patterns=["C3/*"], local_dir="Bio-VQA")

C1 images (TreeOfLife-200M) and the original C2 photographs (CUB-200-2011) are not re-hosted here. See Images for how to get them.

Scoring a model: add a column named <model>_answer with your model's raw replies and run the matching scorer from the GitHub repository. The scorers do all parsing and normalization, the same way for every model.

---

Repository layout

lite/            Bio-VQA-Lite: C1.csv, C2.csv, C3.csv, C4.csv — 500 questions per task, 12,000 in total
C1/              question files, one per organismal group (10 files); images referenced by TreeOfLife-200M uuid
C2/              question file(s); marker-localization renders as zip shards
C3/              one question file and its image shards per task:  C3_<Task>.csv  +  C3_<Task>_NNN.zip
C4/              one question file and its image shards per species × condition:
                 C4_<Species>_<easy|random|hard>.csv  +  C4_<Species>_<condition>_NNN.zip
predictions/     per-item answers of all 18 models and both specialists, exactly as scored in the paper

Shards hold at most 10,000 images each. The large Grevy's zebra pair images use 300 per shard to keep files small.

---

Configurations and splits

Config	Splits	Rows	Notes
lite (default)	C1, C2, C3, C4	12,000	500 per task; the task column names the task
C1_taxonomy	one per organismal group: amphibians … reptiles	890,000	9 question types per image (8 for reptiles)
C2_attributes	test	824,104	10 question families in 4 categories
C3_spatial	counting, relative_counting, absolute_distance, relative_distance, localization, relative_position, viewpoint, ocr	132,504	one split per task
C4_reidentification	easy, random, hard	15,000	5 species × 1,000 pairs per split

Bio-VQA is an evaluation benchmark. All splits are test data, and there is no training split.

---

Components in detail

C1 — Hierarchical taxonomic reasoning

How far does a generalist VLM stand from a taxonomic specialist, and at which rank does it break down?

• Source: a 100K-image subset of TreeOfLife-200M, 10,000 images from each of 10 organismal groups. Six of those groups do not appear in prior biological VQA benchmarks.
• Coverage: 626 families, 2,311 genera, 3,403 species, which is 6× VLM4Bio's genus coverage and 4× its species coverage.
• Sampling constraints: (a) a complete lineage from phylum to species, plus a common name, so no question lacks a ground truth; (b) a minimum image resolution; (c) a per-species cap, so abundant species do not take over a group.

Group	GBIF Backbone filter
Amphibians	class Amphibia
Angiosperms	class Magnoliopsida, Liliopsida
Arachnids	class Arachnida
Birds	class Aves
Bryophytes	phylum Bryophyta
Fish	class Actinopterygii, Teleostei, Elasmobranchii, Holocephali, Coelacanthi, Dipneusti
Gymnosperms	class Pinopsida
Insects	class Insecta
Mammals	class Mammalia
Reptiles	class Squamata, Testudines, Crocodylia, Sphenodontia

Question type	Format	Example answer
phylum	open	Chordata
class	open	Amphibia
order	open	Anura
family	open	Hylidae
genus	open	Dryophytes
species epithet	open	cinereus
scientific name	open	Dryophytes cinereus
common name	open	green treefrog
common name	4-way MCQ	one of four candidates

• Nine groups have 9 questions per image. Reptiles have 8, because the GBIF Backbone moved Squamata to class rank and the order rank is empty. Total: 890,000.
• Recognition vs recall. The common name is asked twice for the same image and target: once open-ended and once as 4-way multiple choice. The three distractors are drawn uniformly from taxa in the same class as the target.
• Common names in English. GBIF vernacular records include Swedish, Danish and Norwegian names. Every non-English common name is resolved to English against iNaturalist and Wikidata.
• Synonyms. A synonym map built from the GBIF Backbone accepts any synonym of the target. For example, a cougar answered as Felis concolor matches Puma concolor.
• Difficulty tiers. At each rank, taxa are split into thirds by image count in TreeOfLife-200M: the top third is Easy, the middle third Medium and the bottom third Hard.

C2 — Fine-grained attribute grounding

Can a VLM find a trait on the animal, and which side of the animal it is on?

• Source: CUB-200-2011, with 11,788 images and 312 binary attributes, each a part · property · value triple (has_wing_color::brown).
• Inversion. Asking 312 sparse yes/no questions would be uninformative, because a constant no scores well. So each part and property are merged into one question (What is the wing color?), and the marked value is the answer. Annotator certainty is used to select and filter candidate traits.
• Distractors are other values attested for the same part and property. Any value that is also true of the image is excluded, so exactly one option is correct and process of elimination does not work.
• Marker localization. A marker is rendered on each image, and the question asks which body part it marks. One distractor shares the marker's color but not its shape, another shares its shape but not its color. The 11,788 renders bring the image count to 23,576.

Category	Family	Questions	Format	Example
Morphological attributes	attr_bare_value	280,433	MCQ	What is the bill color of this bird? → black
Relative salience	attr_group_value	321,073	MCQ	Which … is the most prominently visible physical trait …? → breast color: red
Anatomical localization	marker_localization	11,788	MCQ	What is the body part … marked with the yellow dot? → beak
	visibility_partset	117,880	MCQ	Which of the following parts … are visible? → beak
	visibility_side	35,364	MCQ	Which of the eyes of the bird are visible? → Both
Naming	category_mcq	11,191	MCQ	What type of bird is this? → Gull
	open_common_name	11,788	open	→ California Gull
	open_scientific_name	11,668	open	→ Larus californicus
	open_genus	11,728	open	→ Larus
	open_kind_of	11,191	open	What kind of gull is this? → California
	Total	824,104	~94% MCQ

Answer spaces differ: 15 values for color, 3 for bill length, 4 for visibility_side. CAGS corrects for this per item.

C3 — Quantitative spatial reasoning

Can a VLM count, measure, locate and orient biological subjects?

Following AVA-Bench, spatial reasoning is split into seven atomic tasks plus OCR as a negative control. The source datasets are aggregated into ten 500-image subsets. Count stratification balances sparse and dense scenes, and images whose annotation completeness cannot be verified are dropped. Each task runs only where its ground truth exists.

Task	Items	Runs on	Answer form	Example
Counting	53,498	all ten subsets	integer	How many beetles are there in this image? → 11
Relative counting	42,412	all ten subsets	one of four panels	Which of the 4 images contains the most number of beetles? → Image B
Absolute distance	14,108	2018-NEON-Beetles	number, cm	What is the measurement of the elytra width … with respect to the scalebar …, in cm? → 0.497
Relative distance	3,976	2018-NEON-Beetles	one of four specimens	Which beetle has the largest elytral length? → Beetle D
Localization	3,988	wildlife catalogs	box, four integers	Provide the bounding box coordinate of the zebra in this image. → [854, 819, 3009, 2322]
Relative position	1,000	2018-NEON-Beetles	direction	Where is Beetle (blue) located with respect to Beetle (red)? → Top-right
Viewpoint	13,000	wildlife catalogs	one of ten orientations	What is the viewpoint of the hyena? → left
OCR (control)	522	2018-NEON-Beetles	string	What is the barcode text? → A00000012435
Total	132,504	78,139 images

• Counting tiers. Every canvas has the same pixel budget. Easy is one image filling the canvas. Medium is two panels from different datasets. Hard is two panels from the same dataset.
• Why beetles. Absolute and relative distance need the recorded trait and scalebar measurements, which only 2018-NEON-Beetles has. Relative position needs specimens lying flat at one focal depth, so that top-right is a real two-dimensional relation and not a perspective artifact.
• Why wildlife catalogs. Localization and viewpoint need a single dominant subject with a recorded box and viewing angle. These come from the C4 catalogs (WhaleSharkID, HyenaID, LeopardID, GZCD).
• Localization is evaluated on the Easy tier.

C4 — Individual re-identification

Does a VLM decide by identity (stripes, spots, scars) or by overall resemblance?

Catalog	Species	Identity signal
WhaleSharkID	whale shark	spot pattern
BelugaID	beluga whale	scarring, dorsal ridge (changes between sightings)
HyenaID	spotted hyena	spot pattern
LeopardID	leopard	rosette pattern
GZCD	Grevy's zebra	stripe pattern

Two photographs are placed side by side on one canvas, with one question:

This image shows two images placed side by side. Is the main subject in the left image the very same individual animal as the main subject in the right image?

Difficulty is set with DINOv2 embeddings, retrieved within viewpoint groups:

Condition	Same-individual pair	Different-individual pair	Appearance vs identity
easy	an animal and its nearest match	two animals that look nothing alike	agree
random	drawn without appearance	drawn without appearance	no signal
hard (adversarial)	an animal and its furthest sighting	two nearest neighbors	disagree

Each catalog gives 1,000 pairs per condition, 500 Yes and 500 No, for 15,000 pairs in total. Same-individual pairs are drawn first because repeat sightings are scarce. In the hard condition, the two animals share pose, lighting, background and coat tone, and differ only in the markings that identify them. Pair images are stored under neutral names (img_000001.png, …), because the original file names encode the pair.

---

Bio-VQA-Lite

A stratified, seeded sample of 500 questions per task, for cheap evaluation and for development.

Component	Tasks	Stratified by	Rows
C1	9 question types	organismal group	4,500
C2	4 categories	question family	2,000
C3	8 tasks	source dataset × tier × subset	4,000
C4	3 conditions	species × answer	1,500
Total	24		12,000

Inside each stratum, sampling is uniform with seed 0, and model-answer columns are dropped. Lite rows keep the image reference columns of the full release, so the same load_image helper works. The script is Release/make_lite.py.

Use Lite for development and quick comparisons. Use the full components for numbers you report.

---

Data fields

The columns differ a little between components. These are the ones you need:

Column	Where	Meaning
question	C2, C3, Lite	Question text, exactly as shown to the models
answer	C2, C3, Lite	Gold answer
qa_pairs / qa_pair	C1 / C4	JSON {"Q": question, "A": gold answer}
task	Lite	Task name within the component
uuid	C1	TreeOfLife-200M image identifier
source_url	C1	URL the image was downloaded from
taxonomy columns (kingdom … species, common name)	C1	Full lineage of the image
category	C2	One of the four C2 categories
filename	C2	Path of the original photograph inside the CUB-200-2011 release
source_dataset, tier	C3	Source subset and difficulty / canvas tier, where the task has one
image_zip	C2 (marker rows), C3, C4, Lite	Name of the zip shard that holds the image
image_in_zip	C2 (marker rows), C3, C4, Lite	File name inside the shard
species	Lite C4	Catalog of the pair

import json
q = json.loads(row["qa_pairs"])          # C1;  C4 uses row["qa_pair"]
question, gold = q["Q"], q["A"]

A model's reply should be stored verbatim in a <model>_answer column. The scorers parse it.

---

Images

Component	Images	Where they are
C1	100,000	Not re-hosted. Each row has the TreeOfLife-200M uuid and source_url. Fetch them with Release/fetch_c1_images.py, or match on uuid against TreeOfLife-200M.
C2 originals	11,788	Not re-hosted. Download CUB-200-2011 from the official Caltech release. filename is relative to its images/ root.
C2 marker renders	11,788	Included (C2/*.zip). These are derived images with a drawn marker, and the answer is defined on them.
C3	78,139	Included (C3/C3_<Task>_NNN.zip). Counting canvases and distance overlays are composed by us, and the answer is defined on the composed image.
C4	15,000	Included (C4/C4_<Species>_<condition>_NNN.zip). Side-by-side pairs, stored under neutral names.

Why the split. Images are included wherever the answer is defined on an image we produced: a rendered marker, a composed canvas, a line overlay, a side-by-side pair. These are redistributed under their upstream terms, with per-source attribution. Where a question asks about an unmodified source photograph, we ship the identifier instead.

C1 reproducibility. The C1 download ran concurrently under a per-group budget, so a fresh run of the download script would select a different 10,000 images per group. The released uuid list is the artifact of record.

---

Model predictions

predictions/ holds the per-item raw answers of all 18 models, and of BioCLIP-2 (C1; tol, kingdom and group label sets) and MegaDescriptor (C4, Grevy's zebra), exactly as scored in the paper. Point the scorers in the GitHub repository at these files to regenerate every table in the paper and supplement without running any model.

Models: Qwen2.5-VL-3B/7B, Qwen3-VL-8B/30B-A3B/32B, InternVL3-8B/38B, InternVL3.5-8B/30B-A3B/38B, Gemma-3-12B, Gemma-4-12B, Molmo2-8B, Pixtral-12B, Phi-3.5-Vision, Llama-3.2-11B-Vision, LLaVA-NeXT-Vicuna-7B and LLaVA-NeXT-Mistral-7B. All were run zero-shot with greedy decoding and one attempt per item.

---

How the data were built

Step	C1	C2	C3	C4
Acquire	DuckDB query over the TreeOfLife-200M catalog; concurrent download; verify complete lineage	CUB-200-2011	per-dataset download and preprocessing; ten count-stratified 500-image subsets	catalog download
Clean	non-English common names → English; GBIF synonym map	annotator-certainty filtering	drop images with unverifiable annotation completeness	viewpoint grouping
Generate	8 ranks + MCQ common name	invert part · property · value; render markers	one generator per task; canvas tiers	DINOv2 pair mining (easy / random / hard)
Seeded	✅	✅	✅	✅

The code for every step is in the GitHub repository.

---

Scoring

Component	Graded metric	Floor / ceiling
C1	exact match, soft exact match (GBIF synonyms), Wu–Palmer, Lin, severity	floor: the most frequent taxon at each rank · ceiling: BioCLIP-2
C2	CAGS = (E[s] − E₀[s]) / (1 − E₀[s]), with CIEDE2000 distance for colors and co-occurrence for parts	0 = chance · validated by 3 annotators on 306 items
C3	CAS = max(0, 1 − |n̂ − n| / n) for counting; mean IoU; balanced accuracy; within-20% for distance; lift = (S − φ)/(1 − φ)	image-blind floor φ on every task
C4	accuracy, d′ = z(H) − z(F), criterion c, with log-linear correction	ceiling: MegaDescriptor (Grevy's zebra)

---

Results summary

Eighteen open-weight VLMs from 3.75B to 38.4B parameters, evaluated zero-shot (full table and per-model results in the paper and on GitHub):

	Finding
C1	Exact match falls from .725 at phylum to .123 at scientific name (mean of 18 models). The gap to BioCLIP-2 widens from 1.2× at class (.731 vs .874) to 2× at scientific name (.269 vs .544). Open-ended common names score .039; the same images as 4-way MCQ score .645 (16.4×).
C2	Anatomical localization is the hardest category (.370 CAGS). Given the correct part, 17 of 18 models are at or below chance on left vs right. The exception is pointing-supervised Molmo2-8B.
C3	The best model counts animals exactly on 41.7% of images with 2–10 individuals, against 94.9% on everyday objects in the same range (CountBench). 15 of 18 models fall below an image-blind constant on absolute size measurement, and 12 of 18 fall below a mean-gold-box floor on localization.
C4	MegaDescriptor reaches d′ .733 on Grevy's zebra against a VLM mean of .391. Under adversarial pairing, 17 of 18 models fall below chance (negative d′).

Across components, a strong model is strong everywhere: InternVL3-38B, Qwen3-VL-30B and Qwen3-VL-32B are top five in every component. Below them, the median model moves 5.5 of 18 places between components.

---

Intended use and limitations

Intended use. Zero-shot evaluation of vision–language models on biological images, and diagnosis of which ability fails: rank-level taxonomy, trait grounding, laterality, counting, measurement, localization, orientation, or individual identity.

Not intended for training models that are then evaluated on Bio-VQA. It is also not intended for any operational decision about a real animal or population without expert review.

Known limitations

• Images come from public datasets that models may have seen in pretraining. C1 shares its source with BioCLIP-2's training corpus, so BioCLIP-2 is an optimistic upper bound and not a neutral baseline.
• There is no human performance reference. Human raters validated C2 scoring only, and C1, C3 and C4 ground truth was not audited by humans beyond the source annotations.
• Floors are rule-based; a text-only VLM might set a higher one.
• Coverage is uneven by design: C2 is birds only, distance and relative position are beetles only, and C4 has five species.
• C1 common names are English only. C1 inherits the GBIF Backbone's classification, including its treatment of reptiles and fish.

---

Licenses

Benchmark annotations (questions, answers, Lite, predictions) are released under CC BY-NC 4.0.

Images keep the license of their source dataset:

Source	Used in	Terms
TreeOfLife-200M	C1 (not re-hosted)	per-image licenses recorded in TreeOfLife-200M
CUB-200-2011	C2 (renders included, originals not re-hosted)	non-commercial research and educational use
WAID	C3	per the WAID release
Boxes on Bees and Pollen	C3	per the source release
2018-NEON-Beetles	C3	per the source release
DePondFi	C3	per the source release
GZCD (Grevy's zebra)	C3, C4	per the source release
WhaleSharkID, HyenaID, LeopardID	C3, C4	per the source releases
BelugaID	C4	per the source release

Check the upstream terms before redistributing any image. The code in the GitHub repository is MIT-licensed.

---

Acknowledgments.

This research was supported by the Imageomics Institute, funded by the US National Science Foundation's (NSF) Harnessing the Data Revolution (HDR) program under Award No. 2118240. Qian Ma and Yao Ma are supported by NSF under grant number NSF2406648.

---

Citation

@inproceedings{rayeed2027biovqa,
  title     = {How Finely Can Vision--Language Models See Biology? {Bio-VQA}: A Benchmark for Taxonomic, Morphological, Spatial and Individual Reasoning},
  author    = {Rayeed, S M and Ma, Qian and Stewart, Charles V. and Ma, Yao},
  booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
  year      = {2027}
}

Please also cite the source datasets you use. Their references are listed in the paper and in the GitHub README.

Links: 🌐 Project page · 💻 GitHub · 🤗 This dataset

Contact: open an issue on GitHub or a discussion on this page.