SkillAdam Borrows Adam's Two Moments to Stop Agent Skill Files From Thrashing

arXiv 2609.08944·medium signalXBlueskyLinkedInCopy link

SkillAdam optimizes discrete, non-differentiable skill documents by mapping Adam's structure onto text editing: an optimization memory recording identified problems and prior solution outcomes acts as the first moment to stabilize update direction, while a volatility-driven edit budget tracking history-weighted variation in recent case-level improvements acts as the second moment to control update magnitude. The paper frames the two failure modes as Direction Stability, where corrections get overwritten by iteration-local feedback, and Update Adaptivity, where revision scope should track improvement consistency. Across seven benchmarks spanning short- and long-horizon tasks it reaches state-of-the-art with fewer optimization iterations and lower cost, with code at github.com/ruc-datalab/SkillAdam.