LLM agent errors spread like a contagion: new study quantifies propagation risk

A new paper on arXiv (2607.21912) tackles a pressing question in multi-agent AI: how do erroneous claims propagate through networks of LLM agents? The authors, Ruiwu Niu, Xincheng Shu, and Ying Zhao, formulate a correction-aware network model that tracks agents through susceptible, exposed, infectious, and corrected (SEICS) states. They derive an early-invasion condition for heterogeneous networks and couple it to a majority-vote benchmark where a reliability target imposes a minimum connectivity requirement. Their key theoretical result: under fixed exposure per communication edge, reliability and error control impose opposing graph constraints. They characterize when these constraints have an empty intersection versus an intermediate connectivity range, and identify regular graphs that minimize invasion factors within the reliable class. Under a fixed sender budget, the homogeneous first-order threshold becomes independent of network density, meaning the communication-budget convention determines whether added edges increase early propagation risk.

To validate their theory, the team ran 21,000 simulation trajectories and a controlled experiment with the grok-4.3 model on 36 new closed-world tasks across three six-node topologies. Results show a clear trend: as node degree increased from 2 to 4 and 5, mean first-generation offspring rose from 0.667 to 1.333 and 1.667, while the adoption fraction among exposed neighbors remained constant at 0.333. In the full-cascade subset, mean non-seed erroneous adoption was 0.200, 0.333, and 0.333 respectively. These findings provide a tractable basis for selecting connectivity under explicit reliability and propagation constraints—a crucial step toward designing safer, more robust multi-agent systems in production environments.