Multi-agent debate is a technique in which several instances of a language model each propose an answer and its supporting reasoning, then read one another’s responses and revise their own over multiple rounds until they converge on a common final answer. The idea was set out in “Improving Factuality and Reasoning in Language Models through Multiagent Debate,” posted to arXiv on May 23, 2023 by Yilun Du, Shuang Li, Antonio Torralba, Joshua B. Tenenbaum, and Igor Mordatch.
The mechanism is simple: rather than trusting one chain of reasoning, the system asks several model copies to critique and incorporate each other’s arguments. The authors reported that this substantially improved mathematical and strategic reasoning across tasks and increased factual accuracy, reducing the fallacious answers and hallucinations that single-pass prompting is prone to. The approach works on existing commercial models without any retraining and complements other prompting strategies such as self-consistency and verification. The authors framed it explicitly as a “society of minds,” echoing Marvin Minsky’s older idea that intelligence can emerge from many interacting parts.
The cost is real: running several agents over several rounds multiplies the number of model calls, so the accuracy gains have to be weighed against the extra compute and latency. Debate became one of the standard reference points for the broader argument that orchestrating multiple LLM passes can beat a single larger call, an argument later pushed further by methods like mixture-of-agents.