Consensus-based multi-agent systems may miss valuable disagreement, paper argues

Arxiv CS.CL (NLP/language models)

The paper argues that consensus protocols discard reasoning-trace disagreements that provide a knowledge-representation signal for value-laden tasks. It proposes using disagreement as a resource rather than a problem to resolve.

multi-agent-systemsreasoningknowledge-representation

See 3 more sources

The Deliberative Illusion: Diagnosing Factual Attrition and Stance Homogenization in Multi-Agent LLM Deliberation8 days agoHerun Wan, Jiaying Wu, Minnan Luo, Fanxiao Li, Ningnan Wang, Nancy F. Chen, Min-Yen Kan

Consensus is Strategically Insufficient: Reasoning-Trace Disagreement as a Knowledge-Representation Signal7 days agoMicha{\l} Wawer, Jaros{\l}aw A. Chudziak

the more i use multiple models, the more i think "AI consensus" is a trap — the disagreement is the only part worth paying attention to2 days agowartableappDiscuss

Arxiv CS.CL (NLP/language models)·Sourabrata MukherjeeHamna HamnaKalika BaliSunayana Sitaram

7 days ago