LongMedBench - Visit Cloze: leaderboard
LongMedBench (2026) builds long multi-visit patient histories from MIMIC-IV records to test medical agents on retrieval, temporal reasoning and clinical decisions over long horizons. Temporal reasoning: reinsert masked, timestamped events of one visit into their original positions, scored by mean Kendall tau from -1 to 1; higher is better.
Metric: Mean Kendall tau (-1 to 1; agreement between the ordering given by the model and the true chronological order, averaged over 6,776 items; visit cloze: reinsert masked, timestamped intervention events of one visit into their original positions; MIMIC-IV patients with long multi-visit event streams; naive long-context agent, provider default configuration). Source: arxiv.org. Saturation forecast: Estimated already saturated. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | DeepSeek V3.2 (Thinking) | 0.97 | #406 (DeepSeek V3.2) |
| 2 | GPT-5 Mini | 0.93 | #348 |
| 3 | DeepSeek V3.2 (Non-reasoning) | 0.32 | #406 (DeepSeek V3.2) |
| 4 | Qwen Turbo | 0.05 | #1986 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=longmedbench-visit-cloze · How It Works · Data refreshed daily, snapshot 2026-11-13.