PaperMind - Cross-Source Evidence Reasoning: leaderboard

PaperMind (2026) tests reasoning over real scientific papers from seven domains, with answers rated 1 to 5 by a GPT-4o judge against references taken from the papers. Here the model, as a tool-using agent with the paper PDF, answers questions that need evidence from the paper, its cited works and outside sources; higher is better.

Metric: Judge score for answering research questions that need evidence from the paper PDF, its cited references and external sources, as a smolagents ReAct agent with tools; rated by a GPT-4o judge on a 5-point scale (1-5) against the reference, averaged over the seven scientific domains; higher is better. Source: arxiv.org. Saturation forecast: Around June 2027. 7 models tracked. Updated 14 Nov 2026.

Top models

#ModelScoreOverall rank
1Gemini 2.5 Pro2.99#276
2GPT-4o Mini2.71#1667
3Claude 3 Haiku2.49#2663
4Qwen 3 VL 4B Instruct2.45#1332
5Claude 3.5 Sonnet2.3#752
6Gemma 3 4B (IT)1.81#4139
7Phi-3.5-Vision Instruct1.28#6268

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

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