Research map / Multi-agent systems and games / LLM agents
One Skill, Many Websites: Learning Generalizable Skills Through Polymorphic Abstraction
Simon Yu, Gang Li, Weiyan Shi, Peng Qi
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes PolySkill, a framework that has LLM agents learn reusable web skills via polymorphic abstraction so they generalize across websites rather than overfitting to one site.
Problem
Skill learning for LLM web agents tends to produce skills over-specialized to a single website, which fail to transfer to other sites.
Approach
Abstracts agent-learned skills polymorphically so that a single skill can apply across multiple websites, rather than being tied to one site's specifics. Only the abstract portion is available; details beyond it are unknown.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add polymorphic abstract-class skill structure | ASI-style skill induction (static) · Mind2Web Cross-task | success rate: 52.3 -> 55.4 | Improved |
| add polymorphic structure | skill induction with online update · Mind2Web Cross-task | success rate: 59.4 -> 63.2 | Improved |
| add polymorphic structure | static skill induction · Mind2Web Cross-Domain | success rate: 57.3 -> 60.1 vs baseline 62.3 | No clear effect |
| static skill induction without polymorphism versus no skills | ASI on GPT-4.1 · Mind2Web | success rate: 53.8 -> 52.3 (cross-task) | Hurt |
| add polymorphic structure | skill library size · Mind2Web Cross-task | number of skills: 66 -> 47 (with update) | Improved |
| add online continual update (ablation) | PolySkill · Mind2Web Cross-task | success rate: 55.4 -> 63.2 | Improved |
| train on a single website only (ablation) | skill learner · Target / other sites | skill reusability: below 4% on other sites | Hurt |
| use self-guided exploration instead of fixed sequential curriculum (ablation) | PolySkill · WebArena Shopping | success rate: 42.1 -> 43.1 | Improved |
| use self-guided exploration instead of SkillWeaver curriculum (ablation) | skill induction · WebArena Shopping | success rate: 39.8 -> 43.1 | Improved |
| specialize on new sites after learning on WebArena | ASI vs PolySkill · WebArena Shopping | success rate: +4.9% for PolySkill over ASI | Improved |
| reordering sequential curriculum (ablation) | PolySkill · WebArena Shopping | success rate: 38.2 to 42.1 across orders | Mixed |
Closest papers by meaning
- Beyond Domains: Reusing Web Skills via Transferable Interaction Patterns (NeurIPS 2026 · similarity 0.91)
- Group of Skills: Group-Structured Skill Retrieval for Agent Skill Libraries (NeurIPS 2026 · similarity 0.87)
- SkillOpt: Executive Strategy for Self-Evolving Agent Skills (NeurIPS 2026 · similarity 0.87)
- From Raw Experience to Skill Consumption: A Systematic Study of Model-Generated Agent Skills (NeurIPS 2026 · similarity 0.86)
- SkillOS: Learning Skill Curation for Self-Evolving Agents (NeurIPS 2026 · similarity 0.86)
- Safe and Scalable Web Agent Learning via Recreated Websites (ICML 2026 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
