MTBBench-Multimodal: leaderboard
The multimodal track of MTBBench (Vasilev et al., 2025), a simulated molecular tumor board for head-and-neck cancer: over a multi-turn dialogue an LLM agent requests H&E and IHC slide images, hematology reports and surgical notes for 26 HANCOCK patients and answers 390 GPT-4o-drafted, expert-reviewed true/false and six-option multiple-choice questions, without tools. The score is the unweighted mean of the digital-pathology, hematology and outcome-and-recurrence accuracies in percent; higher is better.
Metric: Accuracy (%; unweighted mean of the three task accuracies; 26 head-and-neck cancer patients from HANCOCK, 390 true/false and six-option multiple-choice questions; multi-turn agent that requests patient files, no tools). Source: arxiv.org. Saturation forecast: Estimated already saturated. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | InternVL3-78B | 69.1 | #733 |
| 2 | GPT-4o (2024-08-06) | 66.7 | #731 |
| 3 | O4 Mini (2025-04-16) | 64.3 | #362 |
| 4 | InternVL3-38B | 63.5 | #905 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=mtbbench-multimodal · How It Works · Data refreshed daily, snapshot 2026-11-12.