SciFig-Bench - Completeness and Correctness: leaderboard
SciFig-Bench (Huang et al., 2026) asks a system to draw the methodology figure of an AI or ML paper from its method text and a figure description, over 435 human-verified papers from 15 venues. Completeness and correctness (0-10): accuracy on 50 yes/no questions per paper about the method's components and information flow, answered from the generated figure alone by Gemini 3.1 Pro, GPT-5.2 and Qwen3-VL and averaged; higher is better.
Metric: Completeness and correctness (0-10) (accuracy on 50 yes/no questions per paper written by Claude Opus 4.6 from the text and author figure, answered from the generated figure alone by Gemini 3.1 Pro, GPT-5.2 and Qwen3-VL and averaged). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 3 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Nano Banana Pro | 7 | |
| 2 | SciFig GPT-5.4-Image (checkpoint unspecified) | 3.7 | |
| 3 | Qwen-Image | 2.2 |
Interactive version: theaggregate.ai/benchmark?slug=scifig-bench-completeness-and-correctness · How It Works · Data refreshed daily, snapshot 2026-11-12.