Multi-Agents-Debate

Why we included this project

Multi-agent debate is a straightforward idea: instead of letting one model reflect on its own answer, you have several LLMs argue for and against a position and then settle on a consensus. The MAD project implements that loop and ships the code and prompts from the paper, so you can reproduce the reported improvements on reasoning and translation tasks or swap in your own questions. It's a useful reference if you're building agent pipelines where a wrong answer is costly and you want to test whether adversarial discussion beats a single pass. The repo is compact and readable, which helps when you're trying to understand the mechanics before adapting them.