PlotGround: A Benchmark for Grounding Plot Digitization in Real Scientific Figures

Forum topic · 小凯 · 2026-10-07

Summary

PlotGround is an automated pipeline for building plot digitization benchmarks from real scientific figures paired with their author-published source data. The pipeline maps figures to source data tables, identifies reconstructable panels, and generates quantitative questions with grounded reference values. Using it, the authors built PlotGround-1k, a human-verified benchmark of 1,119 questions drawn from 1,066 bioRxiv preprints. Across 16 multimodal models, the best model achieves 87.5% accuracy at a ±5% relative error tolerance, but tightening tolerance to ±2% drops every model's accuracy by 11-24 percentage points, revealing a gap between approximate visual reading and precise quantitative recovery. The figure-source pairing also enables a modality comparison: providing the source data table instead of the figure raises coding agents' accuracy from 90.0% to 97.4% while cutting costs by 72%. Paper: arXiv 2610.06825.

Paper Overview

Field: ML Benchmark
Authors: Yaohui Zhang, Binxu Li, Haoyi Duan, Jiacheng Miao, Yixin Wang, Xinran Du, Chenyue Li, Shilong Liu, Kevin Wu, James Zou
Published: 2026-10-05
arXiv: 2610.06825

Abstract (English translation)

Scientific plots often encode quantitative results that are not easily accessible in machine-readable form, making accurate plot digitization important for verifying and reusing published findings. However, it remains unclear how accurately current models can recover plotted values from real scientific figures, since existing benchmarks rely mainly on synthetic plots or cover only a limited range of plot types.

The authors introduce PlotGround, an automated pipeline for building plot digitization benchmarks from real scientific figures together with their author-published source data. PlotGround maps figures to source data tables, identifies reconstructable panels, and generates quantitative questions with reference values grounded in the source data.

Key Results

PlotGround-1k: a human-verified benchmark of 1,119 questions derived from 1,066 bioRxiv preprints.
16 multimodal models evaluated: the best model reaches 87.5% accuracy at a ±5% relative error tolerance.
Tightening the tolerance to ±2% drops every model's accuracy by 11-24 percentage points, exposing a gap between approximate visual reading and precise quantitative recovery.
Figure vs. source data: thanks to the figure-source pairing structure, the same values can be recovered from either modality. Providing the source data table instead of the figure raises coding agents' accuracy from 90.0% to 97.4% while reducing cost by 72%.

Why It Matters

PlotGround provides a reproducible, grounded evaluation of plot digitization on real scientific figures, and its findings quantify how far current multimodal models are from precise quantitative recovery — while also showing that directly supplying source data is both more accurate and cheaper for downstream agents.

---

*Auto-collected on 2026-10-07*