MI-CXR - ICR Variant: leaderboard

ICR Variant questions of MI-CXR (2026), five-option questions over five-visit chest X-ray timelines from MIMIC-CXR, answered zero-shot from all five images: name the type of change between two given visits. Accuracy in percent (chance 20); higher is better.

Metric: Interval-wise Change Reasoning variant in which the pair of visits is specified and only the change type (new, resolved, progressed and similar) is asked, single abnormality, five-option multiple-choice accuracy (%, times 100; chance 20) on MI-CXR, longitudinal reasoning over five-visit chest X-ray timelines from MIMIC-CXR (400 questions in this variant), zero-shot single-step prompting with all five images, temperature 0; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 14 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1GPT-5.276.5#185
2Gemini 3 Pro (Preview)74.3#119
3MedGemma-27B-IT70.5#1350
4Lingshu-32B70.5#1113
5InternVL3.5-8B66.7#1604
6InternVL3.5-14B66.1#1504
7InternVL3.5-38B63.4#1021
8MedGemma-4B61.7#2557
9Claude Sonnet 4.560.1#261
10Qwen 3 VL 32B60.1#622
11Qwen 3 VL 8B59#917
12Lingshu-7B58.5#1635
13IDEFICS2-8B44.8#5013
14MI-CXR DeepSeek-VL-16B (checkpoint unspecified)28.4

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=mi-cxr-icr-variant · How It Works · Data refreshed daily, snapshot 2026-11-21.