DeepTumorVQA - Visual Reasoning: leaderboard
DeepTumorVQA (2026) is a hierarchical visual question answering benchmark on abdominal 3D CT that follows the evidence chain of tumor diagnosis, scored as multiple-choice accuracy on a balanced 10,000-question test subset. This board is accuracy in percent on the visual reasoning questions such as lesion counts and size or laterality comparisons, answered directly from the image input without tools; higher is better.
Metric: Accuracy (%) on the Visual-reasoning-stage questions (two to four options: lesion counts, laterality and size comparisons) of the balanced 10,000-question multiple-choice test subset of DeepTumorVQA (42 subtypes over abdominal 3D CT volumes), direct inference zero-shot with each model's native input (2D organ-centred slices for 2D models, the full 3D volume for 3D models); higher is better. Source: arxiv.org. Saturation forecast: Around August 2028. 14 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Flash | 40.7 | #175 |
| 2 | Meissa-4B | 39.8 | |
| 3 | Qwen 3 VL 4B | 38.8 | #1151 |
| 4 | Qwen 3.5 9B | 38.4 | #828 |
| 5 | Qwen 3.5 35B A3B | 38.2 | #492 |
| 6 | HuatuoGPT-Vision-7B | 36.8 | #2162 |
| 7 | LLaVA-OneVision-7B | 36.6 | #2981 |
| 8 | Qwen 3 VL 8B | 35.6 | #917 |
| 9 | InternVL3-8B | 31.8 | #1782 |
| 10 | LLaVA-Med | 28.7 | |
| 11 | MedGemma 1.5 4B | 26.1 | #2716 |
| 12 | RadFM | 25.4 | |
| 13 | M3D-LaMed-Phi-3-4B | 25.2 | |
| 14 | M3D-LaMed-Llama-2-7B | 23.7 |
No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.
Interactive version: theaggregate.ai/benchmark?slug=deeptumorvqa-visual-reasoning · How It Works · Data refreshed daily, snapshot 2026-11-21.