Survey of Multi-Agent Debate Strategies for LLMs
A systematic literature review characterizes 141 primary studies on Multi-Agent Debate (MAD), a paradigm where multiple LLM-based agents exchange arguments and critique each other's outputs to improve accuracy and robustness. The review derives a three-dimensional taxonomy covering debate participants, interaction mechanisms, and agreement protocols, supported by formal notations. It reveals that the field has converged on a narrow design pattern—static, fully connected topologies, verbatim exchange, short-term memory, and voting resolution—adopted by convention rather than systematic exploration. The study highlights fragmented terminology and a lack of rigorous synthesis of MAD design dimensions.
Key facts
- 141 primary studies on MAD were characterized
- Three-dimensional taxonomy: participants, interaction mechanisms, agreement protocols
- Field converged on static fully connected topologies, verbatim exchange, short-term memory, voting resolution
- Research remains fragmented with inconsistent terminology
- No rigorous synthesis of MAD design dimensions existed prior
- Formal notations render MAD configurations
- MAD aims to improve accuracy and robustness of LLM-based agentic systems
- Convergence on narrow design pattern adopted by convention
Entities
Institutions
- arXiv