New taxonomy reveals multi-agent debates stuck in narrow design patterns

Researchers from the Universitat Politècnica de Catalunya (Motger, Oriol, Marco, Franch) published a comprehensive survey of Multi-Agent Debate (MAD) strategies on arXiv, reviewing 141 primary studies. MAD is a promising paradigm where multiple LLM agents exchange arguments, critique outputs, and iteratively converge on solutions—improving accuracy and robustness. However, the authors found research remains fragmented with inconsistent terminology and no rigorous synthesis of design dimensions.

Their three-dimensional taxonomy categorizes debate participants (who is involved), interaction mechanisms (how they exchange information), and agreement protocols (how they reach consensus). Critically, the analysis reveals the field has converged on a narrow default pattern: static fully connected topologies, verbatim message passing, short-term memory, and majority voting—adopted more by convention than systematic comparison. Promising alternatives (dynamic topologies, abstracted summaries, long-term memory, confidence-weighted resolution) remain underexplored. The authors position the taxonomy as a descriptive map, a framework for controlled benchmarking, and potentially a schema for machine-readable MAD specifications. Future work includes formalizing it into an executable specification for cost-aware benchmarking and automated tuning of debate configurations.