TrueMAD is a new method for multi-agent debate that allows agents to engage in targeted peer advice, improving reasoning and performance in large language models. It improves accuracy over standard MAD by 3.62 and 3.52 percentage points on average.