AutoMedBench - Visual Question Answering: leaderboard
The visual question answering track of AutoMedBench (2026), which asks an LLM agent to complete end-to-end medical-AI research tasks built from public challenge datasets: five radiology, pathology and medical video question-answering tasks. Each run is scored 0 to 100 as the mean of the held-out result (answer accuracy) and a five-stage workflow score (plan, setup, validate, inference, submit), averaged over the Lite and Standard tiers. Higher is better.
Metric: Overall run score (0-100) on the visual question answering track (5 radiology, pathology and video VQA tasks): the mean of the held-out task score (VQA accuracy, scaled to 0-100) and the workflow agentic score (Plan, Setup, Validate, Inference and Submit stages weighted 25/15/35/15/10, LLM judge for the first three); each model is the single agent of the benchmark harness (shared code-execution interface, system prompt, tool schema and default decoding; no vendor agent framework), runs averaged over the Lite and Standard tiers; higher is better. Source: arxiv.org. Saturation forecast: Around January 2028. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GLM-5 | 64 | #260 |
| 2 | Gemini 3.1 Pro (Preview) | 62.3 | #94 |
| 3 | Qwen 3.5 397B A17B | 57 | #246 |
| 4 | MiniMax-M2.5 | 55.8 | #582 |
| 5 | Claude Opus 4.6 | 55.5 | #102 |
| 6 | GPT-5.4 | 36.5 | #123 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=automedbench-visual-question-answering · How It Works · Data refreshed daily, snapshot 2026-11-13.