LLMs in debate systems show increased consensus but also shared errors

A new research paper introduces MAD Bench, a benchmark designed to evaluate how large language models (LLMs) influence each other's decisions in multi-agent debate systems. The study found that LLMs are significantly swayed by peer communication, with debate leading to stronger consensus but also a higher rate of shared errors compared to independent reflection. The research also highlighted that the perceived source of information, rather than its content, can disproportionately affect an LLM's decision-making process. AI

IMPACT Highlights how LLM communication dynamics can lead to both improved accuracy and increased unanimous errors, impacting the reliability of multi-agent AI systems.

RANK_REASON Research paper introducing a new benchmark for LLM behavior in multi-agent systems. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• arXiv
• Hugging Face
• MAD Bench
• multi-agent debate
• Qwen3.5:9b

• paper
• safety

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