AIP: Artificial Intelligence Pedagogy

AIP (Artificial Intelligence Pedagogy) is a research platform for simulating pedagogical interactions between artificial agents using Large Language Models (LLMs). Inspired by Vygotskian theories of social learning, this project explores how structured dialogues between a knowledgeable Teacher LLM and a naïve Learner LLM can facilitate ontology acquisition in a controlled setting.

This work supports the findings presented in the papers:

• Patania et al. (2025) – AI Pedagogy: Dialogic Social Learning for Artificial Agents, accepted at ICSR 2025.
• Patania et al. (2026) – Teaching LLMs Naturally: Pedagogical Strategies for Interactive Knowledge Acquisition, accepted at AAMAS 2026.

---

Project Overview

AIP implements a lightweight “AI Social Gym” where pedagogical strategies are tested through:

• Teacher–Learner dialogues based on different instructional styles (e.g., Top-Down, Bottom-Up, Learner-Driven, Teacher-Guided)
• Ground-truth ontology generation (e.g., artificial alien taxonomies)
• Post-training evaluation using a 20-Questions guessing game

The system uses OpenAI’s GPT-4o via API to simulate both Teacher and Learner agents.

---

Getting Started

Quick Start

For rapid setup, see QUICKSTART.md for a streamlined installation guide.

Installation

For detailed installation instructions, see SETUP.md.

Quick start:

1. Clone the repository
2. Install Python 3.10+
3. Install dependencies:

pip install -r requirements.txt

API Key Configuration

Set your OpenAI API key as an environment variable:

Windows (PowerShell):

$env:OPENAI_API_KEY="your-api-key-here"

macOS/Linux:

export OPENAI_API_KEY=your-api-key-here

Alternatively, create a .env file in the project root.

---

Usage

Single Session

To run a single pedagogical interaction session:

python main.py

Optional arguments:

• --config PATH: Specify a custom config file (default: config.yml)
• --strategy {top_down,bottom_up,learner_questions,teacher_questions,mixed}: Override the strategy
• --skip-training: Skip the training phase
• --skip-testing: Skip the testing phase

Example:

python main.py --strategy top_down --config my_config.yml

Batch Experiments

To run batch experiments across strategies:

python main_batch.py

Expert Testing

To run expert-only testing (without training):

python main_expert.py

Configuration

You can customize the settings in config.yml, including:

• Dialogue strategy (top_down, bottom_up, learner_questions, teacher_questions, etc.)
• Number of trials
• Temperature and max tokens
• Ontology complexity

---

Project Structure

my_aiped/
├── agents/              # Agent implementations (Teacher, Learner, Oracle, etc.)
├── data/
│   └── ontologies/      # Ontology JSON files
├── ontology/            # Ontology generation and utilities
├── prompts/             # Prompt templates
├── testing/             # Testing utilities and 20-Questions game
├── training/            # Training utilities
├── utils/               # Helper functions and API interfaces
├── logs/                # Log files (excluded from git)
├── results/             # Experiment results (excluded from git)
├── visualizations/      # Generated plots and HTML visualizations (excluded from git)
├── main.py              # Main entry point for single sessions
├── main_batch.py        # Batch experiment runner
├── main_expert.py       # Expert-only testing
├── config.yml           # Configuration file
└── requirements.txt     # Python dependencies

---

Experiment Details

Each run consists of:

1. Ontology Generation – A synthetic taxonomy is created with categories like morphology, diet, and habitat.
2. Pedagogical Dialogue – Teacher and Learner agents interact for a fixed number of turns.
3. Evaluation Phase – The Learner plays a 20-Questions game to test ontology acquisition.

Logs of messages and responses are stored in messages.log and responses.log.

---

Testing

Run the test suite with pytest:

pip install pytest pytest-cov
pytest tests/

Run tests with coverage:

pytest tests/ --cov=. --cov-report=html

---

Contributing

This is a research prototype. For contribution guidelines, see CONTRIBUTING.md.

Contributions are welcome in the following areas:

• Bug reports and fixes
• Documentation improvements
• Feature suggestions (aligned with research goals)
• Code quality improvements

---

Citation

If you use this code in your research, please cite:

@inproceedings{patania2025aipedagogy,
  title = {AI Pedagogy: Dialogic Social Learning for Artificial Agents},
  author = {Patania, Sabrina and Annese, Luca and Koyuturk, Cansu and Ruggeri, Azzurra and Ognibene, Dimitri},
  booktitle = {Proceedings of the International Conference on Social Robotics (ICSR)},
  year = {2025}
}

@inproceedings{patania2026teaching,
  title = {Teaching {LLMs} Naturally: Pedagogical Strategies for Interactive Knowledge Acquisition},
  author = {Patania, Sabrina and Annese, Luca and Koyutuerk, Cansu and Ognibene, Dimitri},
  booktitle = {Proceedings of the 25th International Conference on Autonomous Agents and Multiagent Systems (AAMAS)},
  year = {2026},
  publisher = {IFAAMAS},
  address = {Paphos, Cyprus},
  doi = {10.65109/JGLB4831},
  url = {https://doi.org/10.65109/JGLB4831}
}

---

License

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

---

Contact

For questions or collaboration proposals, please contact the corresponding authors:

• Sabrina Patania – sabrina.patania@unimib.it
• Luca Annese – luca.annese1@unimib.it

Alternatively, you may open an issue on GitHub for technical questions or bug reports.

---

Acknowledgments

We acknowledge the contributions of all collaborators and the support of the research community.