Multi-Agent Debate

One model gives you one answer, delivered with the same confidence whether it is right or wrong. Multi-agent debate takes another route: run several instances on the same question, show each of them what the others said, and let them argue it out before anything is returned.

How a round works
The structure is deliberately simple:

• each agent answers independently first, so none of them is anchored to another's opinion;
• each then receives the others' answers with their reasoning, and is asked to critique them and revise its own;
• the loop repeats for a fixed number of rounds and ends on convergence, a majority vote, or a separate judge model's call.

Published work on the method (Du and colleagues, 2023) reported accuracy gains on arithmetic and factual-recall tasks over a single pass.

What it buys and what it costs
Debate spends inference to catch the errors that survive one attempt but not scrutiny — a wrong intermediate step, a half-remembered fact — and the bill is agents multiplied by rounds, in tokens and in latency. Its characteristic failure is conformity: the agents converge politely on a confidently wrong answer and the agreement reads as evidence. Keeping the first round independent, asking explicitly for disagreement, and giving the final call to a judge rather than to consensus all push against that. It differs from self-consistency, which samples many answers and votes but never lets them see one another.