Polyskill Achieves 1.7x Improved Skill Reuse and 9.4% Higher Success Rates through Polymorphic Abstraction in Machine Learning

The challenge of creating truly adaptable artificial intelligence receives a significant boost from new research into skill learning for autonomous agents, spearheaded by Simon Yu, Gang Li, and Weiyan Shi of Northeastern University, along with Peng Qi. These researchers introduce PolySkill, a novel framework that allows agents to learn skills which are not limited to specific websites or tools, but instead generalise effectively across different online environments. Inspired by the concept of polymorphism in software engineering, PolySkill separates the abstract goal of a skill from the specific actions needed to achieve it, allowing for greater flexibility and reuse. Experiments demonstrate that this approach improves skill reuse by a factor of 1. 7 on familiar websites and increases success rates by up to 13. 9% on previously unseen sites, while also reducing the number of steps required to complete tasks, representing a crucial step towards building agents capable of continuous learning in dynamic, real-world settings.

PolySkill Learns Reusable Skills Without Tasks

Researchers have developed PolySkill, a novel approach to artificial intelligence that enables web agents to learn and reuse skills across diverse websites. This work addresses the challenge of creating agents that can effectively operate in dynamic online environments and adapt to changing user interfaces. PolySkill draws inspiration from polymorphic abstraction in software engineering, fundamentally separating a skill’s abstract goal from its concrete implementation. The team implemented this by defining abstract classes that specify high-level goals, and then creating website-specific implementations for those goals.

This separation allows the agent to operate at an abstract level, creating skills that are not tied to the specifics of any single website and are therefore more robust to UI changes. Agents can chain together these abstract operations to execute complex, multi-step tasks. Experiments demonstrate that PolySkill improves skill reuse by 1. 7times on websites the agent has previously encountered, boosts success rates by up to 9. 4% on the Mind2Web benchmark and 13.

9% on previously unseen websites, while simultaneously reducing the number of steps required to complete tasks by over 20%. In self-exploration settings, PolySkill improves the quality of proposed tasks and enables the agent to learn generalizable skills that function consistently across different sites. By allowing the agent to identify and refine its own goals, the framework enhances the agent’s ability to learn an effective curriculum, leading to the acquisition of more transferable skills compared to existing methods. This work provides a practical path toward building agents capable of continual learning in adaptive environments, demonstrating that decoupling a skill’s goal from its execution is a crucial step toward developing autonomous agents that can learn and generalize across the open web.

PolySkill Boosts AI Skill Reuse and Generalization

Scientists have developed PolySkill, a new framework that significantly enhances the ability of artificial intelligence agents to learn and reuse skills while interacting with online environments. This work addresses a key limitation of existing agents, which often struggle to generalize learned skills beyond the specific websites where they were initially trained. The core innovation lies in decoupling a skill’s abstract goal from its concrete implementation, inspired by the concept of polymorphism in software engineering. Experiments demonstrate that PolySkill improves skill reuse by a factor of 1.

7 on websites the agent has previously encountered, boosts success rates by up to 9. 4% on the Mind2Web benchmark and an impressive 13. 9% on previously unseen websites. Notably, the system reduces the number of steps required to complete tasks by over 20%, indicating a substantial increase in efficiency. Researchers confirmed the robustness of these findings by evaluating the framework on both proprietary and open-source agentic models, including Qwen3-Coder and GLM-4.

1. In self-exploration scenarios, PolySkill enhances the quality of proposed tasks and enables the learning of skills that are transferable across different websites. The system achieves this by allowing the agent to identify and refine its own goals, leading to a more effective learning curriculum and the acquisition of more generalizable skills. This research delivers a practical path toward building agents capable of continual learning in adaptive environments, demonstrating that separating a skill’s goal from its execution is a crucial step in developing autonomous agents that can learn and generalize across the open web continuously. The team’s findings suggest that this approach offers a promising direction for developing transferable skills for any agent operating in diverse environments with shared structural patterns.

Polymorphic Skills Boost Agent Adaptability

PolySkill represents a significant advance in the development of web-based artificial intelligence agents, enabling them to learn and reuse skills across diverse online environments. Researchers have introduced a framework inspired by polymorphic abstraction in software engineering, successfully decoupling a skill’s overall goal from the specific methods used to achieve it on individual websites. Experiments demonstrate that this approach improves skill reuse by a factor of 1. 7 on familiar websites and boosts success rates by up to 13. 9% on previously unseen sites, while also reducing the number of steps required to complete tasks by over 20%.

This work addresses a key challenge in artificial intelligence, namely the difficulty of creating agents that can adapt to the constantly changing landscape of the internet. By learning abstract goals rather than website-specific procedures, the agent achieves a level of generalization previously unattainable, successfully mastering both GitHub and GitLab concurrently. The team observed that 73% of learned skills transferred to new websites, a substantial improvement over the 31% achieved by earlier methods. The authors acknowledge that the framework currently faces limitations when dealing with highly dynamic websites where page structures change rapidly, and that the initial quality of abstract skill definitions is crucial for overall performance. Future research directions include extending the framework to handle these dynamic environments, enabling skill sharing between agents, and incorporating human feedback to further refine the learning process. The underlying principle of polymorphic abstraction.