PolySkill: Learning Generalizable Skills Through Polymorphic Abstraction

Official implementation of PolySkill, a framework that enables web agents to learn generalizable and compositional skills through polymorphic abstraction.

Authors: Simon Yu(1), Gang Li(2), Weiyan Shi(1), Peng Qi(2) (1) Northeastern University, (2) Uniphore

Paper: PolySkill: Learning Generalizable Skills Through Polymorphic Abstraction

---

📰 News

[2025-11-10] 🎉 We are excited to release the PolySkill paper and code! Check out our arXiv paper and explore the GitHub repository.

---

🎯 Overview

PolySkill introduces a novel approach to web agent skill learning inspired by polymorphism in software engineering. By separating a skill's abstract goal (what it accomplishes) from its concrete implementation (how it's executed), PolySkill enables agents to:

• Learn skills that generalize across different websites
• Compose complex behaviors from simpler, reusable components
• Improve performance with 1.3-1.8x gains over existing methods
• Self-improve through autonomous exploration

📦 Installation

Prerequisites

• Python 3.9+
• Node.js 18+ (for BrowserGym)
• Chrome/Chromium browser

Quick Install

# Clone the repository (with baseline submodules)
git clone --recursive https://github.com/simonucl/PolySkill.git
cd PolySkill

# Or if already cloned, initialize submodules
git submodule update --init --recursive

# Install dependencies
pip install -r requirements.txt

# Install in development mode
pip install -e .

# Setup BrowserGym
bash setup_browsergym.sh

Environment Variables

Create a .env file with your API keys:

# For OpenAI models
OPENAI_API_KEY=your_openai_key

# For Anthropic models
ANTHROPIC_API_KEY=your_anthropic_key

# Optional: For other open source models in OpenRouter (Qwen3-Coder-480B-A35B, GLM-4.5)
# OPENROUTER_API_KEY=your_openrouter_key
# Skip this if you are hosting your own models via vllm/sglang

Hosting Open-Source Models (Qwen3-Coder-480B-A35B, GLM-4.5)

For running experiments with Qwen3-Coder-480B-A35B and GLM-4.5, you need to host them locally using SGLang before starting experiments:

# Install SGLang
pip install --upgrade pip
pip install uv
uv pip install "sglang" --prerelease=allow

# Start Qwen3-Coder-480B-A35B server (runs on port 30000)
# This requires multiple GPUs (adjust TP_SIZE in the script based on your setup)
./scripts/host_qwen3_coder.sh

# In another terminal, start GLM-4.5 server (runs on port 30001)
./scripts/host_glm4.sh

Important:

• The model servers must be running before starting experiments
• Use screen or tmux to keep servers running in the background
• Adjust TP_SIZE (tensor parallel size) in the scripts based on your GPU count
• Default ports: Qwen3-Coder on 30000, GLM-4.5 on 30001

---

🚀 Quick Start

1. Mind2Web Experiments

Run PolySkill on Mind2Web benchmark:

# Single model, cross-task setting
./examples/run_mind2web.sh gpt-4 cross-task

# All models on all settings
./examples/run_all_mind2web_models.sh cross-task

Available settings:

• cross-task: Generalization to new tasks on seen websites
• cross-website: Generalization to new websites in seen domains
• cross-domain: Generalization to entirely new domains

2. WebArena Experiments

Run PolySkill on WebArena benchmark:

# Single category with GPT-4
./examples/run_webarena.sh gpt-4 shopping

# All models on all categories
./examples/run_all_webarena_models.sh all

Available categories:

• shopping: E-commerce site (187 tasks)
• admin: CMS admin panel (182 tasks)
• reddit: Forum (106 tasks)
• gitlab: Development platform (180 tasks)
• map: OpenStreetMap (109 tasks)
• cross: Cross-website tasks (48 tasks)
• all: All categories

3. Self-Proposing Exploration

Let the agent autonomously explore and learn skills:

# 🛒 Shopping domain (replicates Table 2 from paper)
./examples/run_self_proposing_shopping.sh 150

# 💻 Git platforms (replicates Table 8 from paper)
./examples/run_self_proposing_git.sh 100

---

📊 Reproducing Paper Results

Mind2Web (Table 5, Table 6)

# GPT-4.1 on all settings
python -m polyskill.experiments.mind2web.run_mind2web \
    --config examples/configs/mind2web_polyskill_gpt4.yaml \
    --setting cross-task \
    --model gpt-4.1

# Claude-3.7-Sonnet
python -m polyskill.experiments.mind2web.run_mind2web \
    --config examples/configs/mind2web_polyskill_claude.yaml \
    --setting cross-website \
    --model claude-3.7-sonnet

# Open-source models (Qwen3, GLM-4.5)
python -m polyskill.experiments.mind2web.run_mind2web \
    --config examples/configs/mind2web_polyskill_qwen.yaml \
    --setting cross-domain \
    --model qwen3-coder-480b-a35b

WebArena (Table 7)

# Run all models on shopping category
for model in gpt-4.1 claude-3.7-sonnet qwen3-coder-480b-a35b glm-4.5; do
    python -m polyskill.experiments.webarena.run_webarena \
        --config examples/configs/webarena_polyskill_${model}.yaml \
        --category shopping \
        --model $model
done

Self-Proposing Exploration (Table 2, Table 8)

# 🛒 Shopping domain (Table 2)
python -m polyskill.experiments.self_proposing.run_self_proposing \
    --config examples/configs/self_proposing_shopping_gpt4.yaml \
    --domain shopping \
    --iterations 150 \
    --model gpt-4.1

# 💻 Developer tools (Table 8)
python -m polyskill.experiments.self_proposing.run_self_proposing \
    --config examples/configs/self_proposing_git_gpt4.yaml \
    --domain dev_tools \
    --iterations 100 \
    --model gpt-4.1

---

🔬 Running Baseline Methods

We include ASI and SkillWeaver as git submodules for comparison. To run the baselines:

# Initialize submodules
git submodule update --init --recursive

# Run ASI baseline
cd baselines/ASI
pip install -r requirements.txt
python run_experiment.py --config configs/webarena.yaml

# Run SkillWeaver baseline
cd baselines/SkillWeaver
pip install -r requirements.txt
python main.py --task webarena

For detailed baseline setup and comparison, see baselines/README.md.

Note: PolySkill integrates and extends ASI components in:

• polyskill/core/inducers/asi_inducer.py - ASI-based skill induction with polymorphic extensions
• polyskill/agents/agent/asi_utils/ - Adapted ASI utilities

---

🏗️ Architecture

Core Components

polyskill/
├── core/                       # Skill induction engine
│   ├── core.py                # Main skill induction logic
│   ├── skill_storage.py       # Polymorphic skill storage
│   ├── inducers/              # Skill induction strategies
│   │   ├── pattern_inducer.py  # Pattern-based induction
│   │   ├── polymorphic_inducer.py  # Polymorphic abstraction
│   │   └── asi_inducer.py     # ASI-based induction (adapted)
│   └── judge/                 # Task success verification
├── agents/                    # Agent implementations
│   ├── agent/                 # Base agent architecture
│   │   ├── vlm_based.py       # Vision-language model agent
│   │   ├── llm_based.py       # Language model agent
│   │   └── asi_utils/         # ASI utilities (adapted)
│   └── planner/               # Task planning
├── experiments/               # Experiment scripts
│   ├── mind2web/              # Mind2Web experiments
│   ├── webarena/              # WebArena experiments
│   └── self_proposing/        # Autonomous exploration
├── configs/                   # Configuration files
├── prompts/                   # Prompt templates
├── utils/                     # Utility functions
└── baselines/                 # Baseline methods (git submodules)
    ├── ASI/                   # Agent Skill Induction baseline
    └── SkillWeaver/           # SkillWeaver baseline

Polymorphic Skill Structure

# Abstract class defines the interface
class AbstractShoppingSite:
    def search_product(self, query: str):
        """Searches for a product."""
        pass

    def add_to_cart(self, item_id: str, quantity: int):
        """Adds item to shopping cart."""
        pass

    # Compositional skill
    def find_and_purchase(self, query: str, item_id: str):
        self.search_product(query)
        self.add_to_cart(item_id)
        self.checkout()

# Concrete implementations for specific websites
class AmazonWebsite(AbstractShoppingSite):
    def search_product(self, query: str):
        # Amazon-specific implementation
        click(search_box_id)
        fill(search_box_id, query)
        keyboard_press('Enter')

---

📈 Evaluation Metrics

PolySkill introduces new metrics beyond task success rate:

1. Skill Reusability: Fraction of learned skills used in new tasks
2. Task Coverage: Percentage of tasks that benefit from skills
3. Skill Compositionality: How often skills build upon each other
4. Step Reduction: Average reduction in action steps

View metrics during evaluation:

# Results will show:
# - Task Success Rate: 63.2% (up 9.4% vs baseline)
# - Skill Reusability: 31.0% (vs 18% for ASI)
# - Task Coverage: 58% (vs 45% for baselines)
# - Avg Steps: 4.2 (vs 5.4 for baselines)

---

⚙️ Configuration

Basic Configuration

# examples/configs/my_experiment.yaml

skill_induction:
  enabled: true
  use_polymorphism: true          # Enable polymorphic abstraction
  abstract_class_induction: true  # Learn abstract classes
  composition_enabled: true       # Enable skill composition

  judge_method: webjudge_general
  storage_path: ./my_skills/

model_configs:
  default:
    provider: openai
    name: gpt-4.1
    temperature: 0.1

agents:
  default:
    name: hsm_v3_with_polyskill
    model_config_name: default

Advanced Options

See examples/configs/ for full configuration examples including:

• Custom skill induction strategies
• Multi-model ensemble
• Exploration strategies
• Judge customization

---

📝 Citation

If you use PolySkill in your research, please cite:

@article{yu2025polyskill,
  title={PolySkill: Learning Generalizable Skills Through Polymorphic Abstraction},
  author={Yu, Simon and Li, Gang and Shi, Weiyan and Qi, Peng},
  journal={arXiv preprint arXiv:2510.15863},
  year={2025}
}

---

🤝 Contributing

We welcome contributions! Please see CONTRIBUTING.md for guidelines.

Development Setup

# Install dev dependencies
pip install -r requirements-dev.txt

# Run tests
pytest tests/

# Format code
black polyskill/
flake8 polyskill/

---

📜 License

This project is licensed under the MIT License - see the LICENSE file for details.

---

🙏 Acknowledgments

We acknowledge Orby (now part of Uniphore) for providing the computational resources and infrastructure support that made the experiments in this paper possible. We thank the authors of ASI (Agent Skill Induction) and SkillWeaver for their pioneering work on skill induction in web agents, which inspired and informed this research. We are grateful to SGLang for their exceptional day-0 support and rapid resolution of issues when hosting Qwen3-Coder and GLM-4.5 models. This work builds on BrowserGym for web agent evaluation, and leverages the Mind2Web and WebArena benchmarks for rigorous evaluation.

---

📧 Contact

For questions or issues:

• Simon Yu: yu.chi@northeastern.edu
• GitHub Issues: https://github.com/simonucl/PolySkill/issues

---

🔗 Links

• 📄 Paper: https://arxiv.org/abs/2510.15863
• 💻 Code: https://github.com/simonucl/PolySkill
• 📊 Mind2Web: https://osu-nlp-group.github.io/Mind2Web/
• 🌐 WebArena: https://webarena.dev/

---

Made with care by the PolySkill Team

Back to Top