MedProbeBench (Deep Research Agents) - Claim Recall: leaderboard
Expert-level medical guideline writing by deep research agents on MedProbeBench (2026): each system writes a complete clinical guideline for 50 tumor entities from WHO classification volumes, compared with the expert guideline. Share of the expert atomic claims that the generated guideline covers, times 100; higher is better.
Metric: Task success rate (x100): share of the expert guideline atomic claims that the generated guideline covers (semantic matching, deduplicated), for deep research agents on MedProbeBench: each system writes a complete clinical guideline for 50 WHO-classification tumor entities (one run per task), compared with the expert guideline; higher is better. Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 7 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | MiroThinker-v1.5-pro | 45.2 | |
| 2 | Sonar Deep Research | 44.7 | |
| 3 | O4 Mini Deep Research | 44.4 | |
| 4 | MiroThinker-v1.5 | 43.3 | |
| 5 | Kimi Agent + Kimi-K2.5 | 41.5 | |
| 6 | AgentScope + GPT-4.1 | 40.5 | |
| 7 | Tongyi DeepResearch 30B A3B | 18.9 |
Interactive version: theaggregate.ai/benchmark?slug=medprobebench-deep-research-agents-claim-recall · How It Works · Data refreshed daily, snapshot 2026-11-21.