CharXiv-R: leaderboard
Reasoning split of CharXiv (Princeton, 2024): 1 short-answer question per chart that combines values across plotted elements, over 2,323 arXiv charts, GPT-4o-graded; GPT-4o 47.1%, humans 80.5%.
Metric: Validation chart-visible reasoning accuracy (%). Source: arxiv.org. Source: charxiv.github.io. Saturation forecast: Estimated already saturated. 95 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | O3 (High) | 78.6 | #124 (O3) |
| 2 | O4 Mini (High) | 72 | #180 (O4 Mini) |
| 3 | Claude 3.7 Sonnet | 64.2 | #249 |
| 4 | Claude 3.5 Sonnet | 60.2 | #350 |
| 5 | GPT-4.1 Mini | 56.8 | #355 |
| 6 | GPT-4.1 | 56.7 | #248 |
| 7 | GPT-4.5 | 55.4 | #274 |
| 8 | O1 (High) | 55.1 | #183 (O1) |
| 9 | Doubao-1.5-Pro | 54.4 | #287 |
| 10 | O1 | 52.6 | #183 |
| 11 | InternVL3-38B | 46.4 | #405 |
| 12 | InternVL3-78B | 46 | #352 |
| 13 | Gemini 1.5 Pro | 43.3 | #445 |
| 14 | InternVL3-14B | 43.1 | #505 |
| 15 | Qwen 2 VL 72B | 43 | #460 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=charxiv-r · How It Works · Data refreshed daily, snapshot 2026-10-20.