Daily AI Paper

Today’s standout paper is PlotGround, which tackles a surprisingly important problem: reading real numbers out of scientific figures. In practice, a lot of published results live only in plots, so if you want to verify a claim, reuse data, or build a meta-analysis, you need a model that can recover values accurately from the image itself. The challenge is that synthetic chart benchmarks are too clean, while real figures are messy, diverse, and often only partially reconstructable. PlotGround builds a new benchmark directly from real bioRxiv figures paired with author source tables, so the questions and answers are grounded in the original data. When they test sixteen multimodal models, the best one does well at loose error tolerance, but performance drops sharply when precision matters. That gap is the key takeaway: today’s systems can often estimate a plot, but not reliably recover the exact science behind it.