DeliberationBench: When Do More Voices Hurt? A Controlled Study of Multi-LLM Deliberation Protocols
This is an incremental study that challenges assumptions about the value of complexity in multi-agent AI systems for researchers and practitioners.
The paper tackled the problem of whether multi-LLM deliberation protocols improve performance over simpler methods, finding that a baseline of selecting the best single response outperformed the best deliberation protocol by a 6.0x win rate gap (82.5% vs. 13.8%) with statistical significance.
Multi-agent systems where Large Language Models (LLMs) deliberate to form consensus have gained significant attention, yet their practical value over simpler methods remains under-scrutinized. We introduce DELIBERATIONBENCH, a controlled benchmark evaluating three deliberation protocols against a strong baseline of selecting the best response from a pool of model outputs. Across 270 questions and three independent seeds (810 total evaluations), we find a striking negative result: the best-single baseline achieves an 82.5% +- 3.3% win rate, dramatically outperforming the best deliberation protocol(13.8% +- 2.6%). This 6.0x performance gap is statistically significant (p < 0.01) and comes at 1.5-2.5x higher computational cost. Our findings challenge assumptions that complexity enhances quality in multi-LLM systems.