---
source: https://iinventedlight.com/library/notes/ao-handoff-multiagent-debate-and-moa.md
title: "Multi-agent debate and Mixture-of-Agents"
license: CC-BY-4.0
license_url: https://creativecommons.org/licenses/by/4.0/
source_license: "CC BY 4.0; restated in our own words"
attribution: "Summary by iinventedlight (AI-assisted). Sources: arXiv, [2305.14325] Improving Factuality and Reasoning in Language Models through Multiagent Debate, CC BY 4.0; arXiv, [2406.04692] Mixture-of-Agents Enhances Large Language Model Capabilities, CC BY 4.0."
author: "iinventedlight (synthesis of cited sources); text drafted with AI assistance"
date: 2026-11-04
status: published
checked_on: 2026-11-04
as_of: 2026-11-04
confidence: medium
untested: "No experiment was run; only the fetched arXiv abstract pages were read, so results are as reported by the authors."
category: agent-operations
topic_area: "agent-ops: multi-agent"
tags: agents, multi-agent, debate
sources:
  - id: S1
    title: "[2305.14325] Improving Factuality and Reasoning in Language Models through Multiagent Debate"
    publisher: arXiv
    url: https://arxiv.org/abs/2305.14325
    accessed: 2026-11-04
    license: "CC BY 4.0"
  - id: S2
    title: "[2406.04692] Mixture-of-Agents Enhances Large Language Model Capabilities"
    publisher: arXiv
    url: https://arxiv.org/abs/2406.04692
    accessed: 2026-11-04
    license: "CC BY 4.0"
---
# Multi-agent debate and Mixture-of-Agents

## Short answer
Several model instances that propose answers and debate them over rounds give more accurate maths and reasoning answers and fewer hallucinations. Mixture-of-Agents stacks layers of agents, each reading all outputs of the previous layer, and with open models it beat GPT-4 Omni on AlpacaEval 2.0. Both trade extra calls for quality. [S1][S2]

## Claims
- id: C1
  claim: Multiple instances debate responses over rounds to reach a common answer.
  source: S1
  evidence: "propose and debate their individual responses and reasoning processes over multiple rounds"
- id: C2
  claim: Debate improves factual validity and reduces hallucinations.
  source: S1
  evidence: "reducing fallacious answers and hallucinations"
- id: C3
  claim: Each MoA agent uses all outputs from the previous layer.
  source: S2
  evidence: "Each agent takes all the outputs from agents in the previous layer"

## Sources
- [S1] [2305.14325] Improving Factuality and Reasoning in Language Models through Multiagent Debate. arXiv. https://arxiv.org/abs/2305.14325 . License: CC BY 4.0. Accessed 2026-11-04.
- [S2] [2406.04692] Mixture-of-Agents Enhances Large Language Model Capabilities. arXiv. https://arxiv.org/abs/2406.04692 . License: CC BY 4.0. Accessed 2026-11-04.

License: this note is CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Summary by iinventedlight (AI-assisted).