MMRareBench - Cross-Image Evidence Alignment: leaderboard

MMRareBench (2026) cross-image track: the model describes findings in two or more medical images of different modalities from a rare-disease case, aligns the evidence across them and names their relation. Rubric score (0-100) from a strict Qwen3-VL-235B judge; higher is better.

Metric: Cross-Image Evidence Alignment track (T3) score: per-image findings, their alignment across two or more imaging modalities and the relation type, graded 0-100 by a Qwen3-VL-235B judge with a strict track-specific rubric that awards credit only when a criterion is fully met (five binary dimensions averaged); higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 23 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1GPT-570.8#178
2Gemini 2.5 Pro66.9#276
3Gemini 3 Flash (Preview)62.9#152
4Qwen 3 VL 235B A22B Instruct56.6#532
5Gemini 2.5 Flash54.9#463
6Claude Sonnet 4.554.7#261
7Claude Haiku 4.541.9#583
8Qwen 3 VL 30B A3B Instruct36.9#895
9GLM-4.6V28.7#710
10MedGemma-27B-IT27.3#1350
11Qwen 3 VL 8B Instruct25.7#1029
12Qwen 2.5 VL 32B Instruct21.2#1062
13Qwen 2.5 VL 72B Instruct18.9#841
14GPT-4o15.6#751
15Lingshu-32B8.7#1113

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

Interactive version: theaggregate.ai/benchmark?slug=mmrarebench-cross-image-evidence-alignment · How It Works · Data refreshed daily, snapshot 2026-11-21.