🧠 Multi-Agent Debate for Improving Factuality and Reasoning in Language Models 🤖

Welcome to the Multi-Agent Debate for AI project! In this repository, I explore an innovative approach to improving the factuality and reasoning capabilities of AI agents by engaging them in debates, where they critique and refine each other’s answers. It's like a math quiz 🧮 where five AI students (agents) debate their solutions to come up with the most accurate response.

📜 Paper Summary

This project builds on the concepts introduced in the paper "Improving Factuality and Reasoning in Language Models through Multiagent Debate":

In a Nutshell:

• 🧑‍🏫 Multiple AI Agents Debate: Several AI agents participate in structured debates.
• 🤓 Collaborative Reasoning: Agents critique, refine, and defend their answers, much like a group discussion.
• ✅ Better Accuracy: The collective reasoning helps catch mistakes and refine responses, leading to more accurate and reliable answers.

📚 Takeaways:

• Just like teamwork makes complex problems easier for humans, multiple AI agents debating with each other leads to superior results.
• These debates help improve accuracy and reduce biases by leveraging multiple perspectives.
• Perfect for fields like healthcare, legal research, and scientific analysis where accuracy is crucial.

---

💻 My Project Setup: A Creative Take on AI Debates 🎨

Objective:

I designed a math quiz 🧮 experiment where five AI agents answer questions and engage in debates. Each round brings new critiques and solutions until a winner is declared based on accuracy. The goal? To demonstrate how AI agents improve through peer learning and collaboration.

Technical Setup:

• Model: WizardMath-7B-v1.1 from Hugging Face, fine-tuned for math reasoning.
• Quantization: To make it resource-efficient, I used 4-bit quantization via BitsAndBytesConfig, optimizing the model to run smoothly even on smaller GPUs.
• Dataset: I used the GSM8K dataset, a collection of math word problems, shuffled to randomly select questions for each round.

📊 Visualizing the Debate

Here are some of the key visualizations from the experiment:

• Final Rankings 🥇
• Performance Over Rounds: Watch how each agent's performance fluctuated in the line graph, showing their journey through five rounds of debate.
• Agent Performance in Each Round: Check out the histogram that compares how each agent scored in individual rounds, showcasing the variance in their responses.

🎉 Conclusion

This project is a fun and creative exploration of AI learning through debate! The idea of multiple agents "arguing" over math problems and improving their answers feels like a glimpse into the future of AI collaboration. Hope you enjoy the project as much as I did building it!