Reproduce “PolySkill: Learning Generalizable Skills Through Polymorphic Abstraction” 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
Simon Yu, Gang Li, Weiyan Shi, Peng Qi
PolySkill separates shared operation interfaces from website-specific implementations, allowing higher-level routines to reuse operations across sites. It learns skills from successful browser trajectories, verifies them, and updates its library for new websites. On Mind2Web Cross-Website evaluation with GPT-4.1, online updates reach 61.3% success across 177 tasks, versus ASI’s 58.7%; the full-system comparison does not isolate abstraction’s effect.
Key insights
- Polymorphic abstraction enables the creation of generalizable skills that are reusable across diverse web environments, moving beyond over-specialized approaches.
- Structuring skill learning through a hierarchical, abstract-to-concrete approach accelerates skill acquisition on unseen websites by providing focused exploration goals.
- Effective skill induction, which emphasizes reusability and compositionality, directly correlates with improved task efficiency (fewer steps) and overall task success rates.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.