Alibaba's Qwen Team Unveils SkillDuet, a Co-Evolving Skill-and-Model RL Framework

SkillDuet is a new paper from Alibaba's Qwen team, arXiv 2611.12109, with its main experiments run on Qwen3-4B rather than any model called Qwen4. The approach builds on an existing line of work combining agent skills, reinforcement learning and on-policy distillation, most directly the SEED framework that extracts hindsight skills from trajectories. Its distinct contribution is a two-way loop in which the persistent global skill set and the model parameters evolve together. A new selection criterion called Strength ranks skills by how much they shift the current model's outputs, and only the strongest are merged into the long-term global set. The framework is positioned as a combinatorial innovation within the skill-based agent RL literature rather than an entirely original invention.