Reproduce “MDSKILL: Learning to Internalize Skill Memory through Multi-View Distillation for Efficient Dynamic Agent Self-Evolution” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Dian Jin, Jinluan Yang, Kairong Han, Baohong Li, Zijing Hu, Yi He, Wenlin Zhong, Yunze Tong and 4 more

MDSKILL trains agents to use textual skill memories by separating grounding, planning, and control views and comparing skill-guided with skill-free rollouts from matching task states. It transfers knowledge between views only when rollout outcomes meet its gates. On Qwen3-1.7B, it raises ALFWorld success to 93.8% and WebShop success to 82.8% versus UCOB. WebShop task score nevertheless falls below UCOB.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.