LegendBench: leaderboard
LegendBench (Zhang et al., 2026) tests how vision-language models use chart legends, with generated charts, controlled legend interventions and questions on legend parsing, grounding, reasoning and abstention. QA accuracy is the percentage of all individual questions answered correctly; higher is better.
Metric: QA accuracy (%; proportion of all individual questions answered correctly; LegendBench snapshot of 26,689 legend-centric chart questions generated from 900 base charts with counterfactual legend variants; the same prompt template and hyperparameters for every model). Source: arxiv.org. Saturation forecast: Around February 2027. 12 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.6 Sol | 83.81 | #29 |
| 2 | Qwen 3.5 27B | 71.62 | #358 |
| 3 | Gemini 2.5 Flash | 67.82 | #452 |
| 4 | GLM-4.6V | 64.59 | #698 |
| 5 | GPT-4.1 Mini | 64.19 | #775 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=legendbench · How It Works · Data refreshed daily, snapshot 2026-11-12.