GA-LLM: Hybrid Optimization Framework for Structured Tasks

Genetic Algorithm + LLM + Task-Aware Encoding

Author: Jonas Lin & Jacky Cen
Version: 0.2
License: MIT

---

🔍 What is GA-LLM?

GA-LLM is a modular hybrid optimization framework that combines Genetic Algorithms (GA) with Large Language Models (LLMs) to solve complex, structure-oriented content generation tasks. The system supports tasks like travel itinerary planning, academic proposal writing, and business reporting.

This project is designed to harness the creativity of LLMs and the convergence efficiency of evolutionary strategies, enabling automatic generation, mutation, and selection of structured outputs under constraints.

---

✨ Core Features

• ✅ Modular GA framework with abstract Gene class
• ✅ LLM-driven population generation, mutation, and crossover
• ✅ Prompt templating system with variable injection
• ✅ Multi-level logging with per-run .log files
• ✅ Parallelized scoring and generation with thread pools
• ✅ Constraint validator plug-in architecture
• ✅ Customizable scoring via LLM prompts

---

🗂️ Project Structure

ga-llm/  
├── ga_llm/ # Core GA-LLM framework  
│ ├── base_gene.py # Abstract base class  
│ ├── config.py # Evolution strategy configuration  
│ ├── engine.py # Hybrid evolutionary engine  
│ ├── evaluation.py # LLM-based scoring mechanism  
│ ├── constraints.py # Constraint checker interface  
│ ├── llm_utils.py # LLM calling interface  
│ ├── logging_utils.py # Console + file logger  
│ └── prompts/ # Prompt template manager  
│ └── template_manager.py  
├── demo/ # Task-specific examples  
│ └── travel_demo.py # Travel itinerary optimization  
├── logs/ # Auto-generated log files  
├── requirements.txt  
├── README.md  
└── LICENSE

---

🚀 Getting Started

1. Install dependencies

pip install -r requirements.txt

2. Run the travel itinerary example

python demo/travel_demo.py

You will get an optimized 4-day trip to Shanghai using LLM-guided mutation and crossover.

---

🧠 Sample Output (Excerpt)

Travel Itinerary:
- Days: 4 day program
- Hotels: Shanghai Family Hotel
- Total Cost: ¥4890.00
Detailed Schedule:
[
  {
    "date": "2024-07-01",
    "activities": [
      {"time": "09:00", "location": "Shanghai Museum", "description": "..."}
    ]
  },
  ...
]

---

🧩 Add Your Own Task

To support a new task:

• Implement a new YourTaskGene(BaseGene) class

• Define to_text, parse_from_text, crossover, mutate

• Provide appropriate prompt templates and constraints

• Plug into the framework via HybridEvolutionEngine

---

🤝 Contributing

Contributions are welcome! Please:

• Fork the repo

• Create a feature branch

• Open a pull request with detailed description

We welcome new task types, new scoring logic, prompt improvements, and performance enhancements.

---

📄 License

This project is licensed under the MIT License.
© 2025 Jonas Lin & Jacky Cen