A skill that helps an agent today can become a crutch that prevents it from learning to work unaided. This paper treats skill removal not as a final cleanup step, but as part of the training signal: withdrawing a skill exposes new failures, which can be turned into learning opportunities until the orchestrator relies less on retrieved skills.
AI Agent Papers2026
R$^2$Orch: Recursive Self-Improvement for Agent Orchestration via Self-Play within a Skill-Internalization Loop
Trains an agent orchestrator through recursive self-play, then withdraws skills once they stop improving performance. Learnability-driven task selection and controlled skill evaluations aim to move the system toward deployment without external skills.
Introduction
Information
- Websitearxiv.org
- OrganizationsThe Chinese University of Hong Kong, Shenzhen, China, Fudan University, China, University of Illinois Urbana-Champaign, USA, Tsinghua University, China, National University of Singapore, Singapore
- AuthorsMingda Zhang, Yun Huang, Qiang Huang, Yanjin Li, Rui Qian, Qichao Zhao, Tiesunlong Shen, Na Zhao, Jian Wang, Xiaoying Tang
- Published date2026/10/15
More Items
AI Agent Papers2026
Separates multi-agent prompts into shared coordination rules and agent-specific execution instructions, then guides their optimization according to each role. It can be paired with existing optimizers and improves results across benchmarks in several domains.
AI Video Papers2026
Generates interactive game video from explicit character, scene, and combat state, enabling real-time control and zero-shot generalization to unseen scenes, characters, and actions.
AI Agent Papers2026
Builds executable professional tasks by mapping required information dependencies and repairing gaps revealed during test runs. It produces 112,686 tasks across 612 occupations for collecting teacher demonstrations and training working agents.