Multi-Agent AI Needs Evidence Tracking, Not Just More Agents

Summary

A new arXiv study formalizes the epistemic Sybil problem in multi-agent AI, showing that additional reports may repeat the same evidence or share correlated extraction errors. Experiments with more than 20,000 LLM-agent calls found that increasing reports without increasing evidence roots sharply reduced posterior coverage. The authors recommend tracking evidential ancestry and dependence in collective inference systems.