Homogeneous multi-agent debate is less effective than self-correction

A new research paper, "The Cost of Consensus," reveals that homogeneous multi-agent debate among LLMs is less effective and more costly than isolated self-correction. The study, using models like Qwen2.5-7B and Llama-3.1-8B, found that debate leads to issues like sycophantic conformity, contextual fragility, and consensus collapse. These problems result in debate consuming significantly more tokens for equal or lower accuracy compared to self-correction. AI

Summary written by gemini-2.5-flash-lite from 1 source. How we write summaries →

IMPACT Suggests that current homogeneous multi-agent debate strategies are inefficient and may hinder rather than help LLM problem-solving.

RANK_REASON Academic paper presenting empirical findings on LLM agent behavior. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• Qwen2.5-7B
• Llama-3.1-8B
• Mistral-3-8B
• GSM-Hard
• MMLU-Hard

• paper
• safety