SkillRise: Agentic Reinforcement Learning for Cross-Task Skill Evolution
Zhiyuan Yao, Yuxin Chen, Zhengxi Lu, Zishan Xu et al. · arXiv · 2026
Abstract
Standard agentic reinforcement learning treats every task as an independent episode, so nothing learned on one carries over to the next even when they share obvious structure. This framework orders related tasks into progressively harder sequences and has a single policy alternate between solving and curating a running skill document handed to the next task. Credit is split: solving is judged on the current outcome, curation on discounted downstream ones.
Why it matters
Notable for a test-time result — performance keeps improving as the sequence of related tasks grows longer, with no extra training. That is a different scaling axis from the usual ones.
https://arxiv.org/abs/2607.26784