MedConclusion - Numeric Consistency: leaderboard

MedConclusion (2026) biomedical conclusion generation from PubMed structured abstracts, judged by GPT-5.4-mini for numeric consistency with the author-written conclusion on a 0-100 scale over a random 30K-abstract subset; higher is better.

Metric: Numeric consistency (judge score 0-100): the model writes a formal academic conclusion (prompt A, no length or style constraint) from the non-conclusion sections of a PubMed structured abstract; a GPT-5.4-mini judge scores it 0-100 against the author-written conclusion, averaged over a random 30K-abstract subset of MedConclusion; temperature 0, at most 1,024 new tokens; higher is better. Source: arxiv.org. Saturation forecast: Not forecast. 15 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1GPT-5.4Non-reasoning88.24#135
2Gemini 3.1 Pro (Preview)86.92#99
3Gemini 3 FlashNon-reasoning86.45#175
4DeepSeek V3.2Non-reasoning86.22#419
5Gemma 3 27B (IT)84.13#1324
6Llama 3.2 1B Instruct82.69#9022
7GLM-4.6V80.19#710
8Llama 3.1 8B Instruct79.82#4600
9Qwen 2.5 7B Instruct77.31#2723
10DeepSeek R175.58#482
11Gemma 2 9B (IT)75.05#2369
12MiniMax-M2.173.65#575
13Qwen 2.5 VL 7B Instruct71.82#1902
14Qwen 3 4B 2507 Instruct71.78#2221
15Kimi K2 (Thinking)61.62

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

Interactive version: theaggregate.ai/benchmark?slug=medconclusion-numeric-consistency · How It Works · Data refreshed daily, snapshot 2026-11-21.