Abstracted Web Skills Beat Replaying Raw Trajectories, but Library-Scale Retrieval Gives the Gain Back

arXiv 2609.19523·medium signalXBlueskyLinkedInCopy link

EconSkills (arXiv 2609.19523, submitted 17 Sep 2026) distills verified EconWebArena trajectories into parameterized standard operating procedures, each recording scope, navigation procedure, site-specific guidance, verification checks and recovery steps with source-instance values replaced by placeholders. Separating transfer from retrieval matters: in controlled transfer, matched skills beat no-skill prompting and need fewer steps on paired successes, and abstraction is substantially more effective than replaying raw trajectories. At library scale, however, retrieval is only competitive with the no-skill baseline overall, because approximate matches on uncovered tasks offset the gains on directly covered ones. The design target this implies is coverage-aware selection, not a bigger skill library.