SciDiagramEdit learns to edit scientific figures from arXiv revision histories via agentic skill evolution

arXiv·medium signal

Submitted July 16, 2026 (arXiv 2607.15272) by a team including Ziwei Liu and Jürgen Schmidhuber, this work treats natural paper revisions as a training signal for instruction-driven figure editing on editable vector sources. Its core mechanism is agentic: a proposer 'continually refines the agent's skill specification from execution traces,' progressively improving edit accuracy across validation sets. It's a concrete example of the skill-evolution / self-improving-agent pattern applied to a real research workflow rather than a toy benchmark.