SciImpact: leaderboard

SciImpact (2026) tests whether language models can predict scientific impact: given two papers or artifacts from the same field, pick the one with higher future citations, awards, patent citations, media attention or adoption. The score is the mean pairwise accuracy over seven impact dimensions in percent, chance 50, higher is better.

Metric: Pairwise prediction accuracy (%, times 100) on the seven impact dimensions (unweighted mean of the dimension accuracies), over SciImpact contrastive pairs (215,928 pairs over 19 fields, built from citation counts, best paper awards and Nobel prizes, patent citations, media attention, GitHub stars and Hugging Face downloads with impact thresholds): the model sees two same-field artifacts (title and abstract, README, dataset card or model card, truncated to 1,000 words) and names the one with higher future impact in a two-option forced choice parsed by exact match; the order is balanced so chance is 50%; higher is better. Source: arxiv.org. Saturation forecast: Around November 2028. 11 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1O4 Mini64.2#368
2Claude Haiku 4.563.4#583
3GPT-4.1 Mini62.6#789
4Qwen 2.5 14B59.2#1811
5Qwen 3 4B58.9#2348
6Qwen 2.5 7B58.9#2929
7Llama 3.1 8B56.4#5196
8Llama 3 8B56.2#5900
9Ministral 3 3B55.8#3504
10nemotron-3-nano-30B-a3B54.3#1946
11Llama 3.2 3B53#7877

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

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