Multi-agent debate (MAD) has shown promise in improving reasoning capabilities in large language models, but traditional methods face significant limitations. The reliance on consensus through majority voting introduces randomness and may degrade performance due to conformity among agents. Free-MAD addresses these challenges by implementing a score-based decision mechanism that evaluates debates comprehensively and reduces the number of rounds needed for effective reasoning.
By incorporating anti-conformity, Free-MAD allows agents to resist undue influence from the majority, fostering a more diverse range of opinions and more accurate outcomes. Experiments indicate that Free-MAD not only enhances reasoning accuracy but also significantly lowers token costs, making it more scalable. The framework’s robustness against real-world attacks represents a critical advancement, suggesting its potential for broader applications in AI-driven decision-making processes.
👉 Pročitaj original: arXiv AI Papers
