Conclave – make LLMs debate each other before they respond (adndvlp.github.io)

A new research experiment called Conclave introduces a debate engine for large language models (LLMs), allowing multiple models to discuss and refine their responses before final output. This system leverages existing tools like Claude Code and Gemini CLI, making it accessible to users without extensive modifications to their setup. The debate occurs in structured rounds, with signals such as LEAD, SUPPORT, and CHALLENGE guiding the process. This collaborative approach can surface issues like architecture flaws and security vulnerabilities, enhancing the robustness of the final response. Significantly, Conclave enables users to see real-time reasoning as models deliberate, offering deeper insights into their decision-making processes. The architecture supports mixed teams of LLMs, enabling complex tasks to be tackled through parallel processing and sub-team formations. While the debate approach incurs higher API costs (three calls for a three-model team), it is designed to improve output quality, particularly for intricate assignments. As Conclave is still in its early stages, users may encounter some bugs, but ongoing developments aim to optimize performance and user experience. This innovative approach could transform how the AI/ML community approaches problem-solving and model collaboration.