Anyone who has revised a paper knows the tedium of fixing figures: relabeling parts, shuffling panels, restyling for a reviewer. SciDiagramEdit goes after automating that from a plain-language instruction. The hard part is that a scientific figure isn't a photo, it's a dense infographic where arrows, plots, captions, and schematics follow a tight visual grammar to make an argument, so edits must respect that structure.

Their approach has two pieces. One is a benchmark built by mining before-and-after figure pairs from arXiv version histories, so the edits reflect what authors actually changed. The other is an agent that works on the figure's editable vector source (so you can inspect and tweak individual pieces alongside it) and improves through skill evolution, where a proposer keeps refining the agent's skill description from its own traces. Over several rounds, edit accuracy on held-out figures rises.

This comes only from the abstract, which reports the trend rather than hard numbers, so the paper will have the concrete results.