Bio-VQA - Relative Counting: leaderboard
Fine-grained biological visual QA covering taxonomy, traits, spatial reasoning and individual re-identification. Balanced accuracy fraction. Bio-VQA v1 Table3 native headline columns,18 generalist checkpoints bound to Table13. Greedy decoding, no sampling/chain-of-thought/few-shot; one attempt per item, same prompts and scorer. Do not replace these Table3 measures with differently defined appendix headline metrics. Domain-specialist embedding baselines, Mean and Floor rows are not generic evaluated VLMs and are not admitted.
Metric: Balanced accuracy fraction. Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 18 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | InternVL3-38B-Instruct (Benchmark Sweep Run: source=BioVQA; greedy; zero-shot; noCoT; single-attempt; exact Table13 checkpoint) | 0.57 | |
| 2 | Gemma 4 12B (IT) (Benchmark Sweep Run: source=BioVQA; greedy; zero-shot; noCoT; single-attempt; exact Table13 checkpoint) | 0.56 | |
| 3 | Qwen 3 VL 32B Instruct (Benchmark Sweep Run: source=BioVQA; greedy; zero-shot; noCoT; single-attempt; exact Table13 checkpoint) | 0.56 | |
| 4 | InternVL3.5-38B-Instruct (Benchmark Sweep Run: source=BioVQA; greedy; zero-shot; noCoT; single-attempt; exact Table13 checkpoint) | 0.5 | |
| 5 | InternVL3.5-30B-A3B-Instruct (Benchmark Sweep Run: source=BioVQA; greedy; zero-shot; noCoT; single-attempt; exact Table13 checkpoint) | 0.49 | |
| 6 | Qwen 3 VL 30B A3B Instruct (Benchmark Sweep Run: source=BioVQA; greedy; zero-shot; noCoT; single-attempt; exact Table13 checkpoint) | 0.48 | |
| 7 | Molmo2-8B (Benchmark Sweep Run: source=BioVQA; greedy; zero-shot; noCoT; single-attempt; exact Table13 checkpoint) | 0.42 | |
| 8 | Gemma 3 12B (IT) (Benchmark Sweep Run: source=BioVQA; greedy; zero-shot; noCoT; single-attempt; exact Table13 checkpoint) | 0.42 | |
| 9 | Qwen 3 VL 8B Instruct (Benchmark Sweep Run: source=BioVQA; greedy; zero-shot; noCoT; single-attempt; exact Table13 checkpoint) | 0.41 | |
| 10 | Qwen 2.5 VL 7B Instruct (Benchmark Sweep Run: source=BioVQA; greedy; zero-shot; noCoT; single-attempt; exact Table13 checkpoint) | 0.41 |
Interactive version: theaggregate.ai/benchmark?slug=bio-vqa-relative-counting · How It Works · Data refreshed daily, snapshot 2026-11-06.