PolySkill: Learning Generalizable Skills Through Polymorphic Abstraction

TL;DR

PolySkill tackles the problem of generalizable web-agent skills by introducing polymorphic abstraction, which decouples a skill's abstract goal from its concrete website-specific implementation. The framework defines abstract domain classes (e.g., AbstractShoppingSite) and fills them with concrete site implementations, enabling reusable, composable skills that transfer across websites and domains. A three-stage process combines discovery, refinement, and deployment within a verification pipeline, and new evaluation metrics (Skill Reusability, Task Coverage, Skill Compositionality) quantify transfer and reuse beyond final task success. Across Mind2Web and WebArena benchmarks and multiple foundation models, PolySkill achieves higher generalization, improved efficiency, and stronger continual-learning capabilities, signaling a practical path toward autonomous, continually learning agents on the open web.

Abstract

Large language models (LLMs) are moving beyond static uses and are now powering agents that learn continually during their interaction with external environments. For example, agents can learn reusable skills while navigating web pages or toggling new tools. However, existing methods for skill learning often create skills that are over-specialized to a single website and fail to generalize. We introduce PolySkill, a new framework that enables agents to learn generalizable and compositional skills. The core idea, inspired by polymorphism in software engineering, is to decouple a skill's abstract goal (what it accomplishes) and its concrete implementation (how it is executed). Experiments show that our method (1) improves skill reuse by 1.7x on seen websites and (2) boosts success rates by up to 9.4% on Mind2Web and 13.9% on unseen websites, while reducing steps by over 20%. (3) In self-exploration settings without specified tasks, our framework improves the quality of proposed tasks and enables agents to learn generalizable skills that work across different sites. By enabling the agent to identify and refine its own goals, the PolySkill enhances the agent's ability to learn a better curriculum, leading to the acquisition of more generalizable skills compared to baseline methods. This work provides a practical path toward building agents capable of continual learning in adaptive environments. Our findings show that separating a skill's goal from its execution is a crucial step toward developing autonomous agents that can learn and generalize across the open web continuously.

Figures (6)

• Figure 1: PolySkill, a novel approach that enables web agents to develop polymorphic skills that generalize across websites. PolySkill achieves superior performance with 1.3--1.8$\times$ improvements in task success rate, skill reusability, and compositional capabilities compared to existing methods.
• Figure 2: Limitations of existing skill induction methods. We evaluate ASI and SkillWeaver across two foundation models: GPT-4.1 and Claude-3.7-Sonnet. Both methods show unstable learning dynamics and poor skill reusability, demonstrating over-specialization issues that hurt performance on WebArena Shopping tasks.
• Figure 3: Performance comparison of PolySkill with baseline methods on the Mind2Web benchmark across four large language models. The y-axis shows task success rate (%). The three evaluation settings on the x-axis, Cross-task, Cross-Website, and Cross-Domain, represent increasing levels of generalization difficulty. Our method, PolySkill, consistently outperforms the ASI baseline, while the online continual update version, PolySkill (+Online), achieves state-of-the-art performance across all models and settings. The performance gains are most significant in the challenging Cross-Domain scenario, highlighting our method's superior generalization. Error bars represent the standard error over three runs.
• Figure 4: Overall performance comparison of PolySkill with baselines on the WebArena benchmark across four leading language models. The x-axis shows different website categories, with the leftmost Average group representing the primary overall result. PolySkill consistently achieves the highest average success rate on all models, demonstrating its effectiveness on complex, realistic web tasks. Notably, our method surpasses strong skill-learning baselines like ASI and SkillWeaver, with the most significant gains observed on the powerful GPT-4.1 and Claude-3.7-Sonnet models.
• Figure 5: Relationship between skill reusability and task efficiency in WebArena shopping tasks. Lines show average steps (left y-axis) while bars show Skill Reusability (right y-axis) for ASI (orange), SkillWeaver (green), and PolySkill (red). Higher skill reusability correlates with fewer required steps, demonstrating improved task efficiency through learned skills.