SciDiagramEdit - Evolved Skill Aesthetic: leaderboard
SciDiagramEdit paper v1 test split from 364 author-revision figure pairs. Semantic and aesthetic win rates use blind pairs against the author's revised figure, judged by Claude Opus 4.7. The train/validation/test split is 2:1:3. Native fractions are retained; superscript improvements are not the score. The editor runs through Codex. A skill is an evolved prompt document, not a weight update. The Claude Opus 4.7 Coach uses two epochs, batch size 8, three frontiers, at most three analyst rounds and eight patches per update; editor timeout 600 seconds. Generation effort is not specified. Each backbone evolves its own prompt skill. Metric: Pairwise win rate (0–1).
Metric: Pairwise win rate (0–1). Source: arxiv.org. Saturation forecast: Around January 2027. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 | 0.52 | #31 |
| 2 | GPT-5.4 | 0.44 | #84 |
| 3 | GPT-5.3 | 0.36 | #99 |
| 4 | GPT-5.1 | 0.21 | #140 |
No result here: #5 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=scidiagramedit-evolved-skill-aesthetic · How It Works · Data refreshed daily, snapshot 2026-10-26.