Learning to Compose for Cross-domain Agentic Workflow Generation

Source: arXiv Authors: Jialiang Wang, Shengxiang Xu, Hanmo Liu, Jiachuan Wang, Yuyu Luo, Shimin Di, Min-Ling Zhang, Lei Chen Date: 2026-02-11 Primary category: cs.MA All categories: cs.MA, cs.AI, cs.LG, cs.SE

Abstract

This paper studies cross-domain workflow generation by learning reusable workflow capabilities and composing them in a single pass for new tasks. Instead of relying only on repeated iterative search at deployment time, it decomposes workflows into basis capabilities and learns which combinations actually matter under distribution shift. The result is highly relevant to systems that want workflow evolution without paying a large search tax for every new task.