SciFigQual-Bench: leaderboard
SciFigQual-Bench (University of Sydney and University of Hong Kong, 2026) asks vision-language models to rate published computer-science paper figures for visual clarity, layout, caption consistency, context consistency and misleading risk, reading each figure with its caption and citing paragraphs. Mean absolute error of the overall 1 to 10 score against the mean expert rating on 1,200 figures, one direct call per figure; lower is better.
Metric: Mean absolute error of the overall figure-quality score against the mean expert score (1-10 rating scale, lower is better; 1,200 figures from 254 ACL, EMNLP, ICML and NeurIPS papers, each rated with its caption and citing paragraphs; one direct VLM call per figure). Source: arxiv.org. Saturation forecast: Estimated already saturated. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.8 | 0.44 | #56 |
| 2 | GPT-5.6 Sol | 0.45 | #30 |
| 3 | Gemini 3.5 Flash | 0.46 | #101 |
| 4 | Nova Pro | 0.47 | #1555 |
| 5 | Claude Sonnet 5 | 0.48 | #111 |
| 6 | Llama 4 Maverick | 0.49 | #1099 |
| 7 | Qwen-VL-Max | 0.52 | #1222 |
| 8 | InternVL3-78B | 0.54 | #738 |
| 9 | Seed 2.0 Pro | 0.61 | #211 |
| 10 | GLM-4.6V | 0.66 | #701 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=scifigqual-bench · How It Works · Data refreshed daily, snapshot 2026-11-13.