The impact of multi-agent debate protocols on debate quality: a controlled case study
A novel AI debate system dynamically silences agents to speed up consensus formation.
A new research paper by Ramtin Zargari Marandi isolates a critical variable in AI performance: the debate protocol itself. Multi-agent debate (MAD) systems, where multiple AI models argue to refine an answer, often show gains, but it's been unclear how much comes from the underlying models versus the rules of engagement. This study fixes the models and rigorously tests three protocol designs—Within-Round (WR), Cross-Round (CR), and a novel Rank-Adaptive Cross-Round (RA-CR)—against a no-interaction baseline.
The key finding is that protocol design fundamentally shapes debate dynamics, revealing a trade-off. The Within-Round protocol fostered the highest rate of peer-referencing, while the novel RA-CR protocol achieved the fastest convergence to consensus. RA-CR works by using an external judge model to dynamically reorder the agents and strategically silence the lowest-ranked one each round, efficiently focusing the discussion. For tasks where reaching a unified answer is the priority, this new protocol offers a clear, measurable advantage over existing methods.
- Novel RA-CR protocol uses an external judge to reorder and silence one agent per round, speeding consensus.
- Controlled study on 20 macroeconomic events shows a trade-off: interaction boosts peer-referencing, but RA-CR optimizes for faster convergence.
- Argument Diversity remained high across all main protocols, unaffected by the specific rules of engagement.
Why It Matters
Provides a blueprint for engineers to systematically optimize multi-agent AI systems for speed or collaboration, beyond just upgrading the base models.