Debate Patterns

Three copies of gpt-3.5-turbo that read each other's solutions to a grade-school maths problem and then answer again solve 85 of 100 GSM8K problems, where one copy alone solves 77. That result, published by Du and colleagues in May 2023, made multi-agent debate popular. Give the same number of model calls to a plainer method, nine independent answers and a majority vote, and it reaches 88.2% against the debate's 83.0%, as Huang and colleagues showed five months later. In 2025 Choi and colleagues split debate into its two ingredients, the exchange of arguments and the final vote, and found on seven benchmarks that the vote alone accounts for most of the gain credited to debate.