🏆 AAIPL – Adversarial Language Model Agents

Built using Llama 3.1 for AMD AI Premier League (AAIPL)

---

📌 Overview

This project implements a competitive AI framework where two intelligent agents compete in a head-to-head multiple choice question battle.

The system consists of:

• 🧠 Q-Agent – Generates difficult, domain-specific MCQs
• 🤖 A-Agent – Solves MCQs using structured reasoning
• 📊 Evaluation Module – Measures accuracy and robustness

The entire system was developed using Llama 3.1 within the provided hackathon Jupyter environment.

---

🎯 Problem Statement

In the AMD AI Premier League (AAIPL), teams are required to build:

1. A Question Agent capable of generating challenging and valid MCQs.
2. An Answer Agent capable of solving opponent-generated questions accurately.

The objective is to maximize:

• Question difficulty
• Question correctness
• Answer accuracy
• System reliability

---

🧠 Model Used

🔹 Llama 3.1

• Large Language Model optimized for reasoning
• Used for both Q-Agent and A-Agent
• Controlled via structured prompt engineering
• Tuned for multi-step logical analysis

---

🏗 System Architecture

Jupyter Notebook Environment
        ↓
Llama 3.1 Model
   ├── Q-Agent Prompt Layer
   └── A-Agent Prompt Layer
        ↓
Evaluation Logic
        ↓
Match Results & Accuracy

---

🧠 Q-Agent Design (Question Generator)

The Q-Agent generates domain-specific MCQs using structured prompts.

✅ Features

• Generates 4-option MCQs (A–D)
• Ensures only one correct answer
• Includes multi-step reasoning questions
• Self-verification for logical consistency
• Difficulty-aware question generation

🔎 Strategy

• Conceptual traps
• Edge-case reasoning
• Strict output formatting
• Validation before final output

---

🤖 A-Agent Design (Answer Solver)

The A-Agent solves questions using step-by-step reasoning.

✅ Features

• Chain-of-thought reasoning
• Option-wise logical elimination
• Internal consistency check
• Confidence estimation

📌 Example Output

Final Answer: C  
Confidence: 87%

---

📊 Evaluation Strategy

The system evaluates performance using:

• Accuracy (% correct answers)
• Question validity rate
• Logical consistency check
• Multi-match simulation testing

Multiple simulated matches were conducted to validate robustness.

---

⚙️ Tech Stack

• Python 3.10+
• Llama 3.1
• PyTorch
• HuggingFace Transformers
• Jupyter Notebook (Hackathon Platform)

---

📂 Project Structure

AAIPL_Project.ipynb
│
├── Q-Agent Implementation
├── A-Agent Implementation
├── Evaluation Loop
└── Match Simulation

---

🚀 Key Innovations

• Adversarial co-design of question and answer agents
• Structured prompt engineering for difficulty control
• Self-verification mechanism in Q-Agent
• Reasoning-based validation in A-Agent
• Modular AI architecture within notebook environment

---

🔮 Future Improvements

• LoRA-based domain fine-tuning
• Reinforcement learning for adaptive difficulty
• Multi-domain expansion
• API-based deployment
• Competitive benchmarking system

---

👥 Team

• ML Engineer – Model design & optimization
• Developer – Integration & implementation
• Designer – Documentation & presentation

---

🏁 Conclusion

This project demonstrates how structured prompt engineering combined with Llama 3.1 can create an effective adversarial AI competition framework capable of generating and solving high-difficulty MCQs with strong reasoning and accuracy.

Built entirely within the provided AAIPL Jupyter environment.

---

⭐ If you found this project interesting, feel free to star the repository!