ChartFI - Faithfulness: leaderboard

ChartFI (2026) tests how faithfully and insightfully multimodal LLMs describe complex scientific charts, over 896 charts from peer-reviewed arXiv papers with author-grounded reference descriptions. This board is faithfulness: one minus the share of checkable statements in the description that a Gemini-3.1-Pro judge finds contradicted by the chart, in percent; higher is better.

Metric: Faithfulness (%): one minus the share of verification points in the generated description that a Gemini-3.1-Pro judge finds contradicted by the chart image, times 100, descriptions of the 896 ChartFI-Bench scientific charts (arXiv papers accepted at peer-reviewed venues), temperature 0 and top-p 1; higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 7 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1GPT-5.495.01#135
2Gemini 3 Flash94.71#175
3Claude Sonnet 4.687.12#173
4Qwen 3.5 Plus86.38#242
5Qwen 3.5 27B85.91#375
6InternVL3.5-14B81.34#1504
7Llama 4 Scout73.48#1818

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

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