CLIR-Bench: leaderboard
CLIR-Bench (2026) tests question answering over sparse, irregularly sampled ICU time series from MIMIC-IV: 4-option multiple-choice questions over 11 tasks in understanding, reasoning, forecasting and decision-making, with the full serialized series as input. The board is macro-average accuracy in percent over the 11 tasks; chance is 25 and higher is better.
Metric: Macro-average accuracy (%) over the 11 tasks, 4-option multiple-choice questions over serialized irregular ICU time series (Full-TS input), chance 25; higher is better. Source: arxiv.org. Saturation forecast: Around March 2027. 15 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 Mini | 50.15 | #389 |
| 2 | Kimi K2.6 | 49.09 | #178 |
| 3 | Gemini 2.5 Flash | 40.3 | #453 |
| 4 | DeepSeek V4 Flash | 35.91 | #258 |
| 5 | Gemma 3 27B (IT) | 30.87 | #1289 |
| 6 | Gemma 4 12B (IT) | 29.82 | #693 |
| 7 | Qwen 3.5 9B | 26.64 | #803 |
| 8 | Qwen 3.5 4B | 25.73 | #1175 |
| 9 | Qwen 3.6 27B | 25.01 | #341 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=clir-bench · How It Works · Data refreshed daily, snapshot 2026-11-13.