Multi-Agent AI Systems Must Track Evidential Ancestry, Not Just Report Count

Multi-agent AI systems often improve their inference capabilities by generating numerous reports from various agents and then synthesizing these outputs. However, this paper highlights a critical flaw: simply having more reports does not necessarily mean more independent evidence. The "epistemic Sybil problem" arises when seemingly independent reports actually originate from the same underlying evidence, or when genuinely independent evidence produces highly similar reports, leading to an inflated sense of corroboration. The researchers formalize this problem, demonstrating that aggregators relying solely on report content cannot reliably distinguish true independent corroboration from mere replication. Their Gaussian shared-root model illustrates that common ancestry doesn't imply complete redundancy, but correlated extraction errors, especially from a shared base model, can significantly lower the information ceiling. Extensive testing with LLM agents on synthetic documents confirms these predictions. Naive aggregation of reports from a single evidence root drastically reduces posterior coverage, while increasing the multiplicity of evidence roots restores calibration. The study concludes that collective inference in multi-agent systems must prioritize tracking the ancestry and dependence of evidence, rather than just the quantity or similarity of agent reports, to avoid misleading conclusions.