PlotGround: Grounding Plot Digitization in Real Scientific Figures and Their Source Data
PlotGround tests whether models can recover exact numbers from real scientific plots, not synthetic charts.
The paper introduces an automated pipeline that pairs figures from bioRxiv preprints with author-released source data. Its PlotGround-1k benchmark contains 1,119 human-verified questions from 1,066 preprints. Across sixteen multimodal models, the best result was 87.5% accuracy at a ±5% error tolerance, but all models dropped 11-24 points when tolerance tightened to ±2%. Giving a coding agent the source tables instead of figures raised accuracy from 90.0% to 97.4% and cut cost by 72%. ArXiv · AI/CL/LG's note
The paper introduces an automated pipeline that pairs figures from bioRxiv preprints with author-released source data. Its PlotGround-1k benchmark contains 1,119 human-verified questions from 1,066 preprints. Across sixteen multimodal models, the best result was 87.5% accuracy at a ±5% error tolerance, but all models dropped 11-24 points when tolerance tightened to ±2%. Giving a coding agent the source tables instead of figures raised accuracy from 90.0% to 97.4% and cut cost by 72%. ArXiv · AI/CL/LG's note
score 5