PlotGround: Grounding Plot Digitization in Real Scientific Figures and Their Source Data

Paper & data

arXiv ↗Dataset ↗BibTeX ↓

What is PlotGround?

Plot digitization is the task of reading numbers off a chart: the height of a bar, the median of a box, the y-value of a labeled point. For a model to be useful to scientists, it has to do this on the figures scientists actually publish, which are dense, multi-panel, and messy. But to score a model on a real figure, you need to know the true value behind every mark, and the image alone can’t tell you that.

PlotGround gets that answer from the paper itself. Many preprints deposit the data behind their figures as supplementary tables. PlotGround links a plotted value back to the exact rows that produced it, writes down the operation that turns those rows into the plotted number, and recomputes it. The reference answer comes from the source data, not from someone eyeballing the image.

Here is one of the benchmark’s line-plot questions:

Every question in PlotGround-1k is built this way, across 10 chart types including bar, volcano, scatter, box, line, bubble, strip, and survival plots. Questions name their target only with labels visible in the figure, and models answer from the figure alone.

Leaderboard

Each model sees the full figure and a question targeting one plotted value, and answers from the figure alone. Accuracy is the share of predictions within a relative-error tolerance of the source-grounded reference value. Tightening the tolerance from ±5% to ±2% costs every model between 11 and 24 points.

Model accuracy on PlotGround-1k

16 models from 9 developers, default settings via OpenRouter. Click a column header to sort.

±5%±2%

Rank	Model	Developer	Acc. ±5% ↓	Δ ±5→±2
1	GPT-5.6 Sol	OpenAI	87.5	−13.7
2	GPT-5.6 Luna	OpenAI	83.8	−14.1
3	GPT-5.6 Terra	OpenAI	82.1	−15.5
4	Gemini 3.7 Flash	Google	81.2	−15.5
5	Gemini 3.8 Flash	Google	79.8	−14.3
6	Qwen3.8 Max	Alibaba	77.9	−20.2
7	Claude Opus 5	Anthropic	77.1	−16.9
8	Claude Sonnet 5	Anthropic	74.4	−18.8
9	Gemini 3.1 Pro	Google	67.6	−17.7
10	Qwen3.8 Flash	Alibaba	64.4	−21.8
11	Kimi K3	Moonshot	62.5	−24.3
12	GLM-5.3 Flash	Z.ai	60.2	−20.4
13	Inkling	Thinking Machines	50.3	−18.3
13	MiniMax M3	MiniMax	50.3	−16.5
15	DeepSeek V4.1 Flash	DeepSeek	47.0	−17.7
16	DeepSeek V4 FV exp	DeepSeek	28.6	−11.6

The PlotGround Pipeline

Instead of rendering synthetic charts from data, PlotGround starts from real figures and works backward to ground each plotted value in its source data.

Figure 1. Overview of the PlotGround pipeline: inputs, figure–source mapping, panel-level grounding, and question synthesis.

1. Step 1
   Mapping

   Bind each figure to candidate supplementary tables using direct evidence (explicit references in text) or indirect evidence (figure descriptions matching table contents).

2. Step 2
   Grounding

   Keep only panels whose values can be recovered with simple operations such as selection, aggregation, ratios, or log transforms, and record a reconstruction recipe anchored to literal evidence.

3. Step 3
   Synthesis

   Execute the recipe, sanity-check the recomputed values, and write at most two questions per figure using only figure-visible labels.

Benchmark Examples

Models receive the complete multi-panel figure. Each question names its target only with identifiers visible in the figure; the answer is recomputed from the deposited source table.

BarVolcanoScatterBoxLineBubbleStripSurvival

QuestionIn panel B, what is the bar height for the Menadione concentration of 500 ppm? (y-axis: 'Food consumption (g)')

Source-grounded answer0.098525

Correct if the prediction is within ±5% (or ±2%) relative error of this value. The value is recomputed from the paper's deposited source table, not read from the image.

Figure–Data Discrepancies

During human verification, we found 8 figures whose plotted values disagree with their own source data. GPT-5.6 Sol, the top model on our leaderboard (87.5% at ±5%), read all 8 values as plotted, suggesting that frontier models could help flag such inconsistencies.

Figure 2. GO enrichment plot vs. its deposited source data. The deposited adjusted p-value for vesicle-mediated transport (5.27 × 10⁻⁹) corresponds to −log₁₀ = 8.28, but the plotted bar ends near 7.7, short of the axis limit at 8. Axis clipping cannot explain the gap, since a clipped bar would reach the limit.

Figure 3. Survival curves vs. deposited individual-level records. The published figure shows the blue curve (S/S,+/+) reaching zero survival first, whereas the deposited records imply zero survival for S/S,+/DEL. The genotype labels appear inconsistent with the source data.

Cite PlotGround

If you find PlotGround useful in your research, please cite our work.

@misc{zhang2026plotgroundgroundingplotdigitization,
      title={PlotGround: Grounding Plot Digitization in Real Scientific Figures and Their Source Data}, 
      author={Yaohui Zhang and Binxu Li and Haoyi Duan and Jiacheng Miao and Yixin Wang and Xinran Du and Chenyue Li and Shilong Liu and Kevin Wu and James Zou},
      year={2026},
      eprint={2610.06825},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2610.06825}, 
}