CharXiv-D: leaderboard

Descriptive split of CharXiv (Princeton, 2024): 4 short-answer questions per chart on titles, axis labels, legends and tick values over 2,323 arXiv charts; GPT-4o 84.45%, humans 92.1%.

Source: arxiv.org.

Interactive version: theaggregate.ai/benchmark?slug=charxiv-d · How It Works · Data refreshed daily, snapshot 2026-09-05.