EpiCurveBench: leaderboard
EpiCurveBench (2026) tests chart digitization on 1,000 real-world epidemic curve images from public health reports: a model extracts every plotted series as a table, scored by EpiCurveSimilarity, an alignment metric that tolerates small time shifts and normalizes value errors by the axis range. General-purpose models run at high reasoning effort; scale 0 to 100, higher is better.
Metric: EpiCurveSimilarity, ECS (%): edit distance with real penalty between each extracted and ground-truth series (gap penalty 1, value differences normalized by the y-axis range and counted fully above 1%), scaled to 0-100 and averaged over matched series per chart (0 for a missed series), then over the 1,000 real-world epidemic curve images; the model returns the chart data as a TSV from one generic extraction prompt; the general-purpose VLMs at high reasoning effort, the specialized chart extractors in their single configuration; higher is better. Source: arxiv.org. Saturation forecast: Around June 2027. 6 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Setting | Score | Overall rank |
|---|---|---|---|---|
| 1 | Gemini 2.5 Pro | High | 52.3 | #276 |
| 2 | GPT-5.2 | High | 43.9 | #185 |
| 3 | Claude Opus 4.5 | High | 42 | #160 |
| 4 | Qwen 3 VL 235B A22B Instruct | 27.5 | #532 | |
| 5 | TinyChart | 11.4 | #9868 | |
| 6 | OneChart | 9.4 |
No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.
Interactive version: theaggregate.ai/benchmark?slug=epicurvebench · How It Works · Data refreshed daily, snapshot 2026-11-21.