What's this about?
The paper represents a skill as nodes for execution steps and edges for context-dependent transitions, then evolves populations of these graphs using mutation and crossover.
What exactly is graph-structured skill?
It makes ordering and branching explicit instead of burying a procedure in one prose block. More structure improves searchability but can make flexible tasks harder to encode.
Why does this matter?
Author-reported preprint; 1 HF upvote at fetch time.
- Workflow structure may reduce redundant instructions.
- Population search explores combinations that linear self-editing can miss.
- Reported gains are modest and benchmark-specific.