New ColMAD protocol enhances multi-agent debate for LLMs

A new research paper investigates the effectiveness of multi-agent debate (MAD) in improving large language model (LLM) reasoning. The study finds that existing MAD paradigms, both competitive (CopMAD) and consensus-seeking (CosMAD), suffer from "debate hacking," where agents either produce misleading information to win or filter out disagreements for premature consensus. To address this, the researchers introduce ColMAD, a collaborative protocol that reframes MAD as a non-zero-sum game, encouraging more informative and truthful messages. Experiments show ColMAD outperforms previous MAD protocols by up to 10 percentage points and offers non-trivial improvements over single-agent approaches. AI

IMPACT Proposes a new protocol that could improve LLM reasoning and reduce the need for extensive single-agent fine-tuning.

RANK_REASON Research paper analyzing and proposing improvements to multi-agent debate protocols for LLMs. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• CopMAD
• large-language models
• multi-agent debate
• Yongqiang Chen

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →