Multi-Agent Debate Strategies: Survey, Taxonomy, and Challenges

Multi-Agent Debate (MAD) has emerged as a prominent paradigm for enhancing the accuracy and reliability of Large Language Model (LLM) agents. By allowing multiple agents to critique, argue, and refine their outputs, MAD aims to mitigate common LLM pitfalls like hallucinations and reasoning errors. Despite its popularity, the field has lacked a unified framework, leading to fragmented research and inconsistent terminology.

This paper presents a systematic literature review of 141 primary studies on MAD. The authors establish a three-dimensional taxonomy to categorize how these systems are built: (1) Participant design, (2) Interaction mechanisms, and (3) Agreement protocols. This taxonomy provides a formal language to describe and compare diverse debate architectures.

The authors find that the current research landscape is heavily skewed toward a 'standard' configuration. Most studies utilize static, fully connected networks where agents exchange verbatim text and resolve disagreements through simple voting or consensus mechanisms. While this pattern is popular, the authors argue it is adopted more by convention than by evidence-based comparison. Many promising, more complex design alternatives—such as dynamic topologies or advanced memory-sharing strategies—are currently marginal, leaving the true potential of the MAD paradigm under-investigated.

Because MAD systems involve a complex interplay of design decisions, comparing results across different papers is currently unreliable. This study serves as a foundational map for the field. By formalizing the design space, the authors provide a framework that researchers can use to conduct controlled benchmarking, identify underexplored design choices, and move toward automated tuning of agentic systems. This is a critical step in transitioning MAD from an experimental technique to a robust, engineering-ready methodology.